MoFlo — an opinionated, local-first AI agent orchestration toolkit for Claude Code: semantic memory, learned routing, gates, and spells. No API keys, no cloud, works out of the box.
2026-10-28
ดาว
19
เครื่องมือ
—
อัปไทม์
—
TL;DR
MoFlo makes Claude Code remember what it learns, check what it knows before exploring files, and get smarter over time — all automatically. Install it, run flo init, restart Claude Code, and everything just works: your docs and code are indexed on session start so Claude can search them instantly, gates prevent Claude from wasting tokens on blind exploration, task outcomes feed back into routing so it picks the right agent type next time, and context depletion warnings tell you when to start a fresh session. No configuration, no API keys, no cloud services — it all runs locally on your machine.
Quickstart
npm install --save-dev moflo
flo init
Restart Claude Code. That's it — memory, indexing, gates, and routing are all active.
Or — just ask Claude to install MoFlo into your project and initialize it!
To verify everything is running, run the /healer skill inside Claude Code (or flo healer from the CLI) after restarting — it runs full diagnostics. If anything fails, run /healer --fix to automatically fix issues.
Next Step: Consult the Eldar
After installing moflo, the single highest-leverage thing you can do for the best experience is run /eldar inside a Claude Code session.
Where /healer (or flo healer from the CLI) verifies that moflo itself is wired up correctly, /eldar audits how Claude is set up to actually use your project — guidance docs, CLAUDE.md, memory namespaces, hook/MCP wiring, model routing, and whether every technology in your stack (TypeScript, Python, Rust, Go, etc.) has matching guidance for Claude to lean on. The stack → guidance cross-reference alone is often the difference between "Claude feels lost in this codebase" and "Claude knows this codebase".
/eldar # Read-only audit; categorized findings, severity-ranked
/eldar --fix # Interactive triage — pick what to fix and the Eldar walk you through it
Run it on day one in any new project, any time Claude feels off, or as a periodic health check. Outside of the core install, this is the most impactful thing moflo offers — full details in the /eldar section further down.
Opinionated Defaults
MoFlo makes deliberate choices so you don't have to:
Fully self-contained — No external services, no cloud dependencies, no API keys. Everything runs locally on your machine.
Minimal dependencies — small runtime dep set, all WASM or prebuilt binaries. No native compilation, no node-gyp, no platform-specific build steps.
Node.js runtime — Targets Node.js specifically. All scripts, hooks, and tooling are JavaScript/TypeScript. No Python, no Rust binaries, no native compilation.
node:sqlite (built-in) — The memory database uses Node 22's built-in SQLite engine. No better-sqlite3 native bindings to compile, no WASM round-trip, no platform-specific build steps. Works identically on Windows, macOS, and Linux.
Neural embeddings by default — 384-dimensional embeddings using all-MiniLM-L6-v2. No hash fallback, no peer-optional setup, no install prompts — real semantic search works out of the box. A postinstall step trims the embedding runtime to your platform and strips GPU-only libraries the runtime never loads, reclaiming roughly 340 MB on Linux and 150 MB on Windows from a fresh install. Set MOFLO_NO_PRUNE=1 to skip the trim, or ONNXRUNTIME_NODE_INSTALL_CUDA=true to keep CUDA GPU support.
Full learning stack wired up OOTB — All configured and functional from flo init, no manual setup:
SONA (Self-Optimizing Neural Architecture) — learns from task trajectories
MicroLoRA — fast rank-2 weight adaptations from successful patterns
EWC++ (Elastic Weight Consolidation) — prevents catastrophic forgetting across sessions
HNSW Vector Search — fast nearest-neighbor search over your knowledge base
Semantic Routing — maps tasks to the right agent via learned patterns (ReasoningBank)
Trajectory Persistence — outcomes survive across sessions
All local, no GPU, no API keys, no external services.
Memory-first — Claude must search what it already knows before exploring files. Enforced by hooks, not just instructions.
Task registration before agents — Sub-agents can't spawn until work is tracked. Prevents runaway agent proliferation.
Learned routing — Task outcomes feed back into the routing system automatically. No manual configuration needed — it gets smarter with use.
Incremental indexing — Guidance and code map indexes run on every session start but skip unchanged files. Fast after the first run.
Claude Code is the only target — MoFlo is built and shipped for Anthropic's Claude Code (CLI, IDE extensions, web). It is not a Claude Desktop integration: nothing reads from or writes to ~/.claude/claude_desktop_config.json or %APPDATA%/Claude/, and adding paths there is always a bug. The MCP tools, memory system, and hooks could in principle work with any MCP-capable client, but Claude Code is the only surface we author for, test against, or accept bug reports on.
GitHub-oriented — The /flo skill, PR automation, and issue tracking are built around GitHub. With Claude's help, you can adapt them to your own issue tracker and source control system.
Cross-platform — Works identically on macOS, Linux, and Windows.
Features
Feature
What It Does
Semantic Memory
384-dim domain-aware embeddings. Store knowledge, search it instantly.
Reflection
Distills durable lessons from your sessions into searchable memory — automatically (auto-meditate), or on demand with /meditate.
Cross-Install Sharing
Durable learnings follow you across git worktrees automatically (no setup) — and across your own machines or a whole team via flo memory sync or a git-tracked artifact. Structural namespaces stay local. Full documentation →
Code Navigation
Indexes your codebase structure so Claude can answer "where does X live?" without Glob/Grep.
Guidance Indexing
Chunks your project docs (.claude/guidance/, docs/) and makes them searchable.
Gates
Enforces memory-first and task-creation patterns via Claude Code hooks. Prevents Claude from skipping steps.
Learned Routing
Routes tasks to the right agent type. Learns from outcomes — gets better over time.
Spell Engine
Define multi-step automations as YAML — shell commands, agent spawns, conditionals, loops, memory ops. Full documentation →
/flo Skill
Execute GitHub issues through a full process: research → enhance → implement → test → simplify → PR. (Also available as /fl.)
Spec-Driven Development
/flo -sd runs a spec → plan → implement → verify cycle with memory-indexed artifacts. Opt-in, but once armed both halves are enforced: an implement gate blocks source edits until the plan is reviewed, and verify-before-done gates the PR. Human review is optional (sdd.human_checkpoints, off by default); spec+plan auto-embed into the PR body (embed_in_pr), or point sdd.specs_dir at a tracked path. Details →
tests, test, __tests__, spec, e2e, plus __tests__ dirs inside src/
tests
Languages
Scans detected source dirs for file extensions
.ts, .tsx, .js, .jsx
It also generates:
Generated File
Purpose
moflo.yaml
Project config with detected guidance/code locations
.claude/settings.json
Gate hooks for Claude Code
.claude/skills/flo/
The /flo issue execution skill (also /fl)
CLAUDE.md section
Teaches Claude how to use MoFlo
.gitignore entries
Excludes MoFlo state directories
In interactive mode (flo init without --yes), it shows what it found and lets you confirm or adjust before writing.
Migrating from Claude Flow / Ruflo
If flo init detects an existing .claude/settings.json or .claude-flow/ directory (from a prior Claude Flow or Ruflo installation), it treats the project as already initialized and runs in update mode — merging MoFlo's hooks and configuration into your existing setup without overwriting your data. Specifically:
Hooks — If your .claude/settings.json already has MoFlo-style gate hooks (flo gate), the hooks step is skipped. Otherwise, MoFlo's hooks are written into the file (existing non-MoFlo hooks are not removed).
MCP servers — MoFlo registers itself as the moflo server in .mcp.json. If you had claude-flow or ruflo MCP servers configured previously, those entries remain untouched — you can remove them manually once you've verified MoFlo is working. The /healer skill (or flo healer from the CLI) checks for the moflo server specifically.
Config files — moflo.yaml, CLAUDE.md, and .claude/skills/flo/ follow the same skip-if-exists logic. Use --force to regenerate them.
To force a clean re-initialization over an existing setup:
flo init --force
First-run heads-up: brief CPU spike during initial indexing
The very first time MoFlo runs in your project — typically right after flo init and the next session start — it builds the initial guidance, code map, and test indexes from scratch and generates 384-dim embeddings for every chunk. On a medium-sized project this takes one to a few minutes and you may see elevated CPU during that window. It runs in the background, so you can start working immediately; you just might hear your fans for a bit.
After that first pass, indexing is incremental and lazy — every subsequent session start only re-processes files that actually changed since the last run, which typically finishes in under a second with no perceptible CPU activity. The big one-time cost is a deliberate trade: pay it once, then every future session opens with your project already searchable by meaning.
Tip: Let that initial indexing finish before running /healer (or flo healer). Healer's embeddings and semantic-quality checks expect the indexes to exist — if you run it mid-build it may flag warnings that resolve themselves the moment indexing completes.
2. Review your guidance and code settings
Open moflo.yaml to see what init detected. The two key sections:
Guidance — documentation that helps Claude understand your project (conventions, architecture, domain context):
Tests — test files to index for "what tests cover X?" reverse mapping:
tests:
directories:
- tests # your test files
- __tests__ # jest-style test dirs
patterns: ["*.test.*", "*.spec.*", "*.test-*"]
extensions: [".ts", ".tsx", ".js", ".jsx"]
exclude: [node_modules, coverage, dist]
namespace: tests
MoFlo chunks your guidance files into semantic embeddings, indexes your code structure, and maps test files back to their source targets — so Claude searches your knowledge base before touching any files. Adjust these directories to match your project:
flo memory index-guidance # Index your guidance docs
flo memory code-map # Index your code structure
flo healer # Verify everything works (alias: flo doctor)
Both indexes run automatically at session start after this, so you only need to run them manually on first setup or after major structural changes. The first index may take a minute or two on large codebases (1,000+ files) but runs in the background — you can start working immediately. Subsequent indexes are incremental and typically finish in under a second. To reindex everything at once:
MoFlo automatically indexes three types of content on every session start, so your AI assistant always has up-to-date knowledge without manual intervention.
What gets indexed
Index
Content
What it produces
Namespace
Guidance
Markdown files in your guidance directories (.claude/guidance/, docs/, etc.)
Chunked text with 384-dim semantic embeddings — enables natural-language search across your project documentation
guidance
Code map
Source files in your code directories (src/, packages/, etc.)
Structural index of exports, classes, functions, and types — enables "where does X live?" navigation without Glob/Grep
code-map
Tests
Test files matching configured patterns (*.test.*, *.spec.*)
Reverse mapping from test files to their source targets — enables "what tests cover X?" lookups
tests
How it works
Session start hook — When Claude Code starts a new session, MoFlo's SessionStart hook launches the indexers sequentially in a single background process. This runs silently — you can start working immediately.
Incremental — Each indexer tracks file modification times. Only files that changed since the last index run are re-processed. The first run on a large codebase may take a minute or two; subsequent runs typically finish in under a second.
Embedding generation — Guidance chunks are embedded using MiniLM-L6-v2 (384 dimensions, WASM). These vectors are stored in the SQLite memory database and used for semantic search.
No blocking — The indexers run in the background and don't block your session from starting. You can begin working immediately.
Configuration
Each indexer can be toggled independently in moflo.yaml:
auto_index:
guidance: true # Index docs on session start
code_map: true # Index code structure on session start
tests: true # Index test files on session start
Set any to false to disable that indexer. The underlying data remains in memory — you just stop refreshing it automatically. You can still run indexers manually:
Without auto-indexing, Claude Code starts every session with a blank slate — it doesn't know what documentation exists, where code lives, or what tests cover which files. It resorts to Glob/Grep exploration, which burns tokens and context window on rediscovery.
With auto-indexing, Claude can search semantically ("how does auth work?") and get relevant documentation chunks ranked by similarity, or ask "where is the user model defined?" and get a direct answer from the code map — all without touching the filesystem.
Learning From Your Sessions
MoFlo turns the work you do into durable knowledge. A lesson worth keeping — a reusable pattern, a gotcha that cost you an hour, a decision and the rationale behind it — lands in the learnings memory namespace, where it's embedded and semantically searchable in every future session. Two paths feed that namespace: one you trigger, one that runs on its own.
/meditate — deliberate retrospective
Run /meditate at the end of a meaningful chunk of work. It reviews the session, distills the handful of lessons that would help a future session on a different task, deduplicates them against what's already stored, and writes the survivors to learnings. It's the curated pass — you choose when to capture, and it keeps only high-signal items (an empty reflection is a valid result, not a failure to pad).
/meditate # Review the whole session and store durable lessons
/meditate --preview # Show the candidate lessons without writing anything
/meditate <focus> # Scope the retrospective to one thread, e.g. "the auth refactor"
Auto-meditate — the automatic counterpart
Auto-meditate does the same job without being asked. While you work, a lightweight hook notices when a durable lesson emerges — a correction, an error you fixed, a decision you made. At the next session start, a brief background pass distills those into learnings using the same durability bar and dedup-then-store protocol as /meditate. It reuses your existing Claude Code session (no extra API key), runs in the background, and ships on by default.
auto_meditate:
enabled: true # Set to false to opt out — /meditate still works manually
The two are complementary: auto-meditate is the always-on safety net, /meditate is the deliberate, curated pass. Both write to the same learnings namespace and both deduplicate, so neither pollutes the other — and anything they store is available to semantic search from then on.
Sharing learnings across installations
The learnings (and knowledge) namespaces are the durable slice of memory — user-authored and worth carrying across checkouts. The rest of the DB (code-map, guidance chunks, run state) is structural or ephemeral and rebuilds cheaply, so it stays local. For ongoing sync MoFlo shares only the durable slice — never live-shares the whole moflo.db between two daemons, which would make each daemon's in-memory search index diverge.
Learnings converge across worktrees at session-start (stored at <git-common-dir>/moflo/durable.db). Override with memory.durable_path; opt out with memory.worktree_sharing: false
Your own laptop + desktop + CI
flo memory sync --to/--from <file>
Export to a synced folder (Dropbox/iCloud) or a copied file, import on the other machine
Teammates' session-start import-merges it after git pull (newest edit wins; one file per entry so concurrent PRs don't conflict; author provenance retained)
A fresh workspace that must be ready fast
flo memory backup/restore or memory.hydrate_from
Seed a new workspace from a whole-DB snapshot so it's searchable on session one — no cold reindex. Safe because each workspace owns its own copy (a one-time restore, not live sharing)
Verify your setup with flo doctor -c shared-db (it warns if you've accidentally pointed at a full moflo.db for live sharing). Full guide: moflo-cross-install-memory-sharing.md.
The Gate System
MoFlo installs Claude Code hooks that run on every tool call. Together, these gates create a feedback loop that prevents Claude from wasting tokens on blind exploration and ensures it builds on prior knowledge.
Gates explained
Gate
What it enforces
When it triggers
Why it matters
Memory-first
Claude must search the memory database before using Glob, Grep, or Read on guidance files
Before every Glob/Grep call, and before Read calls targeting .claude/guidance/
Prevents the AI from re-exploring files it (or a previous session) already indexed. Forces it to check what it knows first, saving tokens and context window.
TaskCreate-first
Claude must call TaskCreate before spawning sub-agents via the Task tool
Before every Task (agent spawn) call
Ensures every piece of delegated work is tracked. Prevents runaway agent proliferation where Claude spawns agents without a clear plan.
Task-status(blocks by default)
The task list Claude opened must reflect reality before the PR — every task closed, deleted, or explicitly acknowledged as deferred
Before gh pr create, whenever the session opened tasks and left some open
A stale task list is worse than none: it reports as not-started work that is about to merge. Every other step here is gated on an artifact; this is the one that gates what the user is shown. Deliberate deferrals stay legitimate — the block prints a one-command acknowledgement. Set task_status_gate: warn to report without blocking, or off; task_create_first: false disables both halves.
Context tracking
Tracks conversation length and warns about context depletion
On every user prompt (UserPromptSubmit hook)
As conversations grow, AI quality degrades. MoFlo tracks interaction count and assigns a bracket (FRESH → MODERATE → DEPLETED → CRITICAL), advising Claude to checkpoint progress or start a fresh session before quality drops.
Routing
Analyzes each prompt and recommends the optimal agent type and model tier
On every user prompt (UserPromptSubmit hook)
Saves cost by suggesting haiku for simple tasks, sonnet for moderate ones, opus for complex reasoning — without you having to think about model selection.
Verify-before-done(on by default)
Claude must verify the change end-to-end (the /verify skill) before gh pr create
Before gh pr create (docs-only diffs exempt)
Enforces "prove it works before done." On by default since #1294 (/flo delegates to /verify and reuses its test run — no double verify). Opt out with verify_before_done: false or --no-verify. Pairs with the Spec-Driven Development cycle, which gives verification its acceptance criteria.
Implement gate (SDD front half)
When a run is armed for SDD, Claude must author a spec and get its plan reviewed before editing source
Before every source Write/Edit, when -sd/sdd.default armed the run
Makes the SDD front half enforced, not advisory (#1297): a spec-gated run can't skip to implementation. No-op for non-SDD work. Opt out per-project with sdd_gate: false, per-run with --no-sdd. Pairs with verify-before-done to gate both ends of the cycle.
Smart classification
The memory-first gate doesn't blindly block every request. It classifies each prompt:
Simple directives (e.g., "commit", "yes", "continue", "looks good") — skip the gate entirely, no memory search required
Task-oriented prompts (e.g., "fix the auth bug", "add pagination to the API") — gate enforced, must search memory first
Escape hatch
Prefix any prompt with @@ to bypass the memory-first gate for that turn. Useful for conversational questions, thinking out loud, or discussions that don't need prior context:
@@ what do you think about this approach?
@@ question — is there a better way to handle auth tokens?
The @@ prefix is stripped before Claude sees the prompt, so it won't affect the response.
Disabling gates
All gates are configurable in moflo.yaml:
gates:
memory_first: true # Set to false to disable memory-first enforcement
task_create_first: true # Set to false to disable TaskCreate enforcement (disables task_status_gate too)
task_status_gate: block # Open tasks block `gh pr create`; warn → report only; off → silent
context_tracking: true # Set to false to disable the context-usage notice
# context_limit: 200k # Optional. Your model's window, so the notice can quote a %.
# Unset (default) = report raw tokens; moflo never guesses a window.
verify_before_done: true # On by default (#1294); set false to skip /verify before `gh pr create`
You can also disable individual hooks in .claude/settings.json by removing the corresponding hook entries.
The /flo Skill
Inside Claude Code, the /flo (or /fl) slash command drives GitHub issue execution:
/flo <issue> # Full process (research → implement → test → PR)
/flo -t <issue> # Ticket only (research and update ticket, then stop)
/flo -r <issue> # Research only (analyze issue, output findings)
/flo -s <issue> # Swarm mode (multi-agent coordination)
/flo -h <issue> # Hive-mind mode (consensus-based coordination)
/flo -n <issue> # Normal mode (default, single agent, no swarm)
/flo -w <issue> # Worktree: run the work in a fresh git worktree (aliases: -wt, --worktree)
/flo -sd <issue> # SDD: spec → plan → implement → verify (implies --verify)
/flo -v <issue> # Verify-before-done only (no spec/plan front-half)
/flo -m <issue> # Auto-merge the PR once required checks pass
Flags compose: e.g. /flo -sd -m <issue> runs the SDD cycle and auto-merges. Each modifier has a --no-* form (--no-sdd, --no-verify, --no-merge) to override a moflo.yaml default for a single run — including --no-verify, since verify-before-done is on by default. For full options and details, type /flo with no arguments — Claude Code will display the complete skill documentation. Also available as /fl.
Provisioned worktrees (-w / flo worktree)
Running two tickets at once means two worktrees, and a bare git worktree add gives you a valid
checkout that you cannot actually run: no node_modules, none of your gitignored .env files, and
dev servers that fight the primary checkout over the same ports. /flo -w drives flo worktree add,
which creates the worktree and provisions it.
Provisioning is opt-in per project. With no worktree: block in moflo.yaml, flo worktree add
creates the worktree and provisions nothing — exactly what a plain git worktree add would do.
worktree:
dir: ../myrepo-worktrees # default: <repo-parent>/<repo>-worktrees
copy: [".env", ".env.*"] # gitignored files copied from the primary checkout
link: ["node_modules"] # symlinked (junctioned on Windows) from the primary checkout
setup: "npm ci" # run inside the new worktree after copy/link
Key
Reach for it when
Watch out for
copy
A fresh checkout is missing gitignored config your build needs
It relocates secrets outside the repo, and outside its .gitignore. Sources must live inside the primary checkout — ../secrets is refused.
link
node_modules is large and you are not using npm workspaces
A symlinked root node_modules is fragile under npm/yarn workspaces — prefer setup: npm ci there. An existing path is never clobbered.
setup
Install or build steps must run per workspace
A non-zero exit marks the provision failed but leaves the worktree in place.
Ports. MoFlo cannot rewrite your hardcoded ports — it has no way to know which files hold them.
Instead each worktree gets a small integer, unique among live worktrees and reused when one is
removed, exported to the setup command as MOFLO_WORKTREE_INDEX. Offset your own ports from it:
worktree:
setup: "npm ci && node -e \"require('fs').writeFileSync('.env.local','PORT='+(3000+Number(process.env.MOFLO_WORKTREE_INDEX)*20))\""
flo worktree remove refuses a tree with uncommitted changes unless you pass --force, and names
any gitignored files it discarded that provisioning did not create. MoFlo's own
.moflo/worktree.json never counts as "uncommitted work" — otherwise every worktree it created
would demand --force — but un-pushed specs under .moflo/specs/ do.
flo worktree is a plain CLI command with no MCP wrapper, so it works the same whether you type it
in a terminal or Claude runs it through Bash. /flo -w takes the second path — that is all the flag
does.
One thing to know if you drive it yourself from a session, and the reason /flo -w handles it for
you: Claude Code resets the Bash working directory to the project root after every call. Two
consequences, both silent when you get them wrong:
# ✗ the cd is gone by the next call — this resolves the WRONG repo
cd path/to/repo
flo worktree add feature/42-thing --json
# ✓ bind the directory to the call
cd path/to/repo && flo worktree add feature/42-thing --json
# ✗ runs the tests in the PRIMARY checkout; the run goes green and the PR is empty
cd "$WORKTREE"
npm test
# ✓ bind it, or skip cd entirely where the tool supports it
cd "$WORKTREE" && npm test
git -C "$WORKTREE" status
flo worktree resolves the repository from its working directory, so a call made from the wrong
place targets a different repository — usually reported as Not a registered worktree of this repo,
occasionally as work landing somewhere you did not intend. When editing files in the worktree, use
absolute paths under the path add printed rather than repo-relative ones.
Read the worktree location from --json rather than reconstructing it; the sibling-directory
convention lives in one tested function, and a second copy of that rule will drift from it.
If you are on a flo older than the skills synced into .claude/ — which happens in a MoFlo source
checkout before a change is published, or when a global flo lags the project's devDependency —
the command reports Unknown command: worktree. /flo -w falls back to a plain git worktree add
and tells you provisioning was skipped, rather than failing the run.
Spec-Driven Development (SDD)
/flo can run the full spec → plan → (review) → implement → verify cycle — the 2026 agentic-coding pattern — with two independent modifiers. Turning SDD on is opt-in, but once a run is armed for it (via -sd or sdd.default) both halves are enforced: the front half by an implement gate, the back half by verify-before-done.
-sd / --sdd(opt-in; enforced once on) — author a spec (the what + acceptance criteria) and a plan (the steps) before implementing. When armed, the check-before-implement gate blocks every source Write/Edit until the spec exists and its plan is reviewed (#1297) — a run can't skip straight to code. Artifacts persist as Markdown at <specs_dir>/<slug>/{spec,plan}.md (default .moflo/specs) and are indexed into memory, so prior specs are searchable across sessions. Implies --verify.
-v / --verify(on by default) — the verify half: a normal run plus the verify-before-done gate. It runs the /verify skill, which exercises the change end-to-end against its acceptance criteria and reports a per-criterion PASS/FAIL, reusing the run's own tests (no double verify). Runs by default; --no-verify skips it. Separable from SDD, so you get "prove it works" without the spec ceremony.
Human review is optional and off by default. moflo stays autonomous: the front half self-advances its spec/plan review checkpoints without stopping. Set sdd.human_checkpoints: true to sit in the loop and approve each artifact by hand before the cycle proceeds.
Both compose with execution mode (-n/-s/-h) and --worktree, and default from moflo.yaml:
sdd:
default: false # true → every /flo run uses the SDD cycle unless --no-sdd
specs_dir: .moflo/specs # where spec/plan artifacts are written (see below)
human_checkpoints: false # true → pause for human approval at spec & plan review (default: autonomous)
embed_in_pr: true # true → append spec+plan to the PR body so reasoning is reviewable
gates:
verify_before_done: true # on by default (#1294); false → skip /verify unless -v. Per-run: --no-verify
sdd_gate: true # on by default (#1297); false → don't enforce the spec→plan front half
On upgrade, consumers with no verify_before_done key start enforcing verify-before-done; an explicit value is preserved. sdd_gate only bites when a run is actually armed for SDD, so non-SDD work is unaffected. Docs-only diffs are exempt, so a pure-docs PR is never blocked.
Where specs live — and making them reviewable. By default spec/plan artifacts are written under .moflo/specs, which flo init gitignores — so they stay local and never bloat source control — and (with embed_in_pr, the default) the spec + plan are appended to the PR body, so the reasoning is reviewable in the PR even while the files stay local. If you'd rather source-control the artifacts, point sdd.specs_dir at a tracked path (e.g. docs/specs or .specs) and commit them normally; no gitignore surgery required — the two mechanisms compose. The path is written with / and resolved cross-platform. If you set it to a directory already covered by guidance.directories, moflo indexes the specs once (as guidance) rather than twice.
--sdd implies --verify (a spec/plan without an enforced verify step drifts). Manage artifacts directly with flo sdd (spec, plan, review, check, embed, list), and start from a fuzzy idea with /commune, which can hand its synthesized spec straight into the SDD spine. Wiring status is reported by /healer (and /eldar).
Auto-merge
-m / --merge — after the PR is opened, await its merge preconditions and merge it, instead of stopping at "PR opened". Orthogonal to execution mode (-n/-s/-h), --worktree, and --sdd/--verify, and always runs after the quality gates (tests, simplify, learnings, verify) have let gh pr create through — so --merge never bypasses a gate.
It prefers GitHub-native auto-merge, falls back to poll-then-merge once required checks are green, and — when the only remaining blocker is review-required on a repo you administer (e.g. a solo repo where GitHub blocks self-approval) — auto-attempts an admin squash-merge. If an unattended admin merge is denied by Claude Code's permission classifier (headless/cron runs), it hands you a copy-paste command rather than looping. It re-confirms checks are green before any admin merge, and never --admin over a red or pending required check.
Default seeds from moflo.yaml; the per-run flag overrides, and --no-merge opts a single run out:
merge:
auto: false # true → every /flo run auto-merges its PR unless --no-merge
Epic handling
When you pass an issue number, /flo automatically checks if it's an epic — no extra flag needed. An issue is treated as an epic if any of these are true:
It has a label matching epic, tracking, parent, or umbrella (case-insensitive)
Its body contains a ## Stories or ## Tasks section
Its body has checklist-linked issues: - [ ] #101
Its body has numbered issue references: 1. #101
The issue has GitHub sub-issues (via the API)
When an epic is detected, /flo processes each child story sequentially — full process per story (research → implement → test → PR), one at a time, in the order listed.
For simple epics with independent stories, /flo <epic> is all you need. For complex features where you want state tracking, resume capability, and auto-merge between stories, use flo epic instead.
Feature Orchestration (flo epic)
flo epic is the robust epic runner — it adds persistent state, resume from failure, and per-story auto-merge on top of /flo. It takes a GitHub epic issue number:
flo epic 42 # Fetch epic #42, run all stories sequentially
flo epic 42 --dry-run # Preview execution plan without running
flo epic 42 --strategy auto-merge # Per-story PRs with auto-merge between stories
flo epic status 42 # Check progress (which stories passed/failed)
flo epic reset 42 # Reset state for re-run
flo epic fetches the epic from GitHub, extracts child stories from checklists, numbered references, and ## Stories / ## Tasks sections, then runs each through /flo with state tracking. If a story fails, you can fix the issue and re-run flo epic 42 — it resumes from where it left off, skipping already-passed stories. (flo epic run 42 is an explicit alias for the same shorthand.)
/flo <epic>
flo epic <epic>
State tracking
No
Yes (epic-state memory namespace)
Resume from failure
No
Yes (skips passed stories)
Auto-merge PRs
No
Yes (--strategy auto-merge)
Dry-run preview
No
Yes
Spells
What are spells?
Spells are declarative YAML automations composed of pluggable step commands. They exist because shell scripts drift, ad-hoc prompts aren't reproducible, and CI/CD pipelines are the wrong tool for local automation. A spell is deterministic (same inputs → same steps), reviewable (a YAML file you read like a recipe), and replayable (re-cast it tomorrow and it behaves the same). Spells run from the CLI (flo spell cast), from an MCP tool call inside Claude Code, or on a schedule.
How do they work?
Each cast goes through the same lifecycle:
Parse & validate — YAML is parsed and every step's config is checked against its command's schema.
Resolve capabilities — every step declares what it needs (shell, net, fs:read, fs:write, memory, credentials, browser, agent). Undeclared access is rejected before execution.
Execute the step graph — steps run sequentially by default, or in parallel groups, with depends_on for ordering and condition/loop for flow control.
Outputs flow forward — any step can reference a prior step's output as {stepId.field}, so later steps can consume earlier results.
Persistence — memory writes, artifacts, and per-step logs persist between runs; pause/resume and dry-run are supported.
flo spell cast -n development # Cast a named spell
flo spell cast -f ./my-spell.yaml # Cast from a file
flo spell cast -n sa --dry-run # Validate without casting
flo spell list # List available spells and recent runs
flo spell grimoire list # Browse built-in spell templates
flo spell schedule list # List scheduled spells
Defining spells
name: my-spell
steps:
- name: lint
command: npm run lint
- name: test
command: npm test
- name: deploy
command: ./deploy.sh
depends_on: [lint, test]
Steps support shell (bash), agent spawns, memory reads/writes, conditionals, loops, parallel groups, browser automation, GitHub/IMAP/Outlook/Slack/MCP integrations, prompts, waits, graphs, and composite steps. See docs/SPELLS.md for the full schema.
Sandboxing and security
Spells run with least-privilege access. Each step command declares the capabilities it needs (shell, net, fs:read, fs:write, memory, credentials, browser, browser:evaluate, agent), and the runner blocks any undeclared access. Spell authors can further restrict capabilities per step — e.g. fs:read: ["./config/"] or shell: ["cat", "jq"] — but never expand them.
Bash steps also run inside an OS sandbox when one is available. MoFlo auto-selects the best tier installed on your machine:
Tier
Platform
Isolation
Docker
all
Container, strictest
bwrap
Linux
User namespaces
sandbox-exec
macOS
Apple seatbelt profile
none
fallback
Capability-only enforcement
The sandbox is network-off by default; a step must explicitly declare net to reach the outside world. Credentials referenced as {credentials.X} are resolved from the encrypted credential store, masked in logs, and never written to disk. A destructive-pattern checker refuses to run bash commands that look like rm -rf /, unscoped git push --force, and similar footguns. See docs/SPELL-SANDBOXING.md for the full model.
Reusability
You interact with spells at three tiers:
Shipped spells — bundled with MoFlo and ready to cast. Browse them with flo spell grimoire list.
User spells — YAML files you drop in .claude/spells/. Override the location in moflo.yaml:
A user spell with the same name as a shipped one wins — that's how you customize a shipped spell without forking.
Custom step commands and connectors — drop TypeScript/JavaScript files in .claude/spells/steps/ (new step types) and .claude/spells/connectors/ (new connectors). They're auto-discovered at startup. Connectors that wrap heavy SDKs (IMAP, MCP) declare those SDKs as optionalDependencies; install them only if your spells use them.
Scheduling
Spells can run on a schedule. The MoFlo background daemon polls for due spells once a minute and casts them — no external cron, no extra services.
Daemon prerequisite. Schedules only fire while the daemon is running. To survive reboot:
flo daemon install # registers an OS-level autostart service
flo daemon status # shows whether the service is registered AND running
flo spell schedule create warns when the daemon isn't installed so you don't quietly miss runs.
Monitoring.The Luminarium — moflo's localhost daemon dashboard — surfaces live schedules, recent executions, and per-schedule controls (disable / re-enable / run now), alongside worker health, memory stats, and Claude Code session stats. Ask /luminarium in your Claude session and it'll print the link.
For full configuration (scheduler: block in moflo.yaml), event types, and the catch-up window after restarts, see docs/SPELLS.md#scheduling.
Building spells with the /spell-builder skill
Inside Claude Code, use the /spell-builder skill to create, edit, and validate spell definitions interactively. The skill understands the full spell schema and available step commands, so you can describe what you want in natural language and it will generate the YAML:
/spell-builder # Start the spell builder
Other AI-client skills shipped with MoFlo
Beyond /flo, /spell-builder, and /eldar, MoFlo ships a handful of focused slash-command skills that work in any consumer project once you flo init:
Skill
Purpose
/guidance
Author and audit guidance docs. Default writes guidance for Claude into .claude/guidance/ as Markdown applying moflo's universal rules. -h switches the audience to humans (lighter ruleset, writes into docs/). --html emits HTML with a minimal default stylesheet instead of Markdown. -a audits every doc in .claude/guidance/ against the universal rules.
/flo-simplify (alias /distill)
Adaptive code review on the current diff. Tier-based fan-out — trivial edits get a self-review, small diffs get one routed agent, cross-cutting refactors get three parallel agents. Routes through the moflo model router for cost-aware execution. (Named /flo-simplify to avoid colliding with Claude Code's built-in /simplify; also available as /distill for a more wizardy feel.)
/quicken (alias /perf-audit)
Ad-hoc performance audit — "quicken the codebase". Reveals N+1 queries, needless re-renders, missing caching, memory leaks, and redundant computation, then prescribes the fix inline in the session. Scopes to your current diff by default (or a path you name), so it never re-audits the whole tree. Replaces the old always-on optimize daemon worker.
/ward (alias /test-gaps)
Ad-hoc test-gap audit — "ward the codebase". Reveals untested functions, uncovered edge cases, and missing error-handling/integration tests, then conjures ready-to-paste test skeletons in your framework inline in the session. Scopes to your current diff by default (or a path you name). Replaces the old always-on testgaps daemon worker.
/commune
Socratic requirements elicitation. Turns a fuzzy "I'm not sure exactly what I want yet" idea into a concrete spec through a short Q&A, then hands the result off to a /flo ticket, a spell, or memory. The pre-execution counterpart to /meditate — use it to open a unit of work.
/divine
Multi-hop web research with explicit confidence gating. Plans the inquiry, searches the web, scores its own confidence, and keeps digging until the answer is well-supported (or a hop cap is hit) — then returns a cited synthesis and remembers what worked so the next research run starts smarter.
/meditate
Deliberate session retrospective — distills durable lessons into the learnings memory namespace, deduped against what's already there. See Learning From Your Sessions for the full picture, including its automatic counterpart, auto-meditate.
/spell-schedule
Schedule a spell on the local moflo daemon (cron, interval, or one-time) without leaving the chat. For remote Anthropic-cloud agents on a schedule, use Claude Code's built-in /schedule instead.
Run any of them with no arguments to see full usage, or browse the source in .claude/skills/ (each skill is a single SKILL.md file).
Epics
Epics are a specialized process for handling GitHub issues that contain multiple child stories. When you pass a GitHub issue to /flo and it's detected as an epic, MoFlo processes each child story sequentially through the full /flo process (research → implement → test → PR).
For simple epics, /flo <epic-number> is all you need. For complex features requiring state tracking, resume from failure, and auto-merge between stories, use the dedicated flo epic command:
flo epic 42 # Run all stories in epic #42
flo epic 42 --strategy auto-merge # Per-story PRs with auto-merge
flo epic 42 --dry-run # Preview execution plan
flo epic status 42 # Check progress
flo epic reset 42 # Reset state for re-run
See the Epic handling section above for detection criteria and the comparison between /flo <epic> and flo epic run.
The Luminarium
The Luminarium is moflo's localhost daemon dashboard. It boots automatically with the background daemon (no extra service to install) and stays running as long as the daemon is up.
Finding the URL
Inside a Claude Code session in a moflo project, ask /luminarium — the skill prints the dashboard URL for the current project. flo daemon status prints the same URL alongside the health summary. Each project binds its own port so two projects on the same machine never collide.
What it shows
Tab
What you see
Workers
Live agent processes the daemon is running (statusline updater, indexer, embedder, etc.)
Schedules
All registered spell schedules (cron / interval / one-time), with run-now and disable controls
Memory namespace breakdown, vector count, embedder backend, HNSW index health
Claude Stats
Per-session Claude Code transcript stats — tokens (with a subagent-token subtotal that splits main-loop context from Task-tool-spawned subagent context), tools called, files touched (local primary sessions only)
Flags
flo daemon start --no-dashboard — disable the HTTP server entirely (the daemon itself still runs)
flo daemon start --dashboard-port <N> — pin to a specific port, overriding the per-project default. Also accepts the MOFLO_DAEMON_PORT env var, which the rest of moflo respects when talking to the daemon
Commands
You don't need to run these for normal use — flo init sets everything up, and the hooks handle memory, routing, and learning automatically. These commands are here for manual setup, debugging, and tweaking.
Memory
flo memory store -k "key" --value "data" # Store with 384-dim embedding
flo memory search -q "auth patterns" # Semantic search
flo memory index-guidance # Index guidance docs
flo memory code-map # Index code structure
flo memory rebuild-index # Regenerate all embeddings
flo memory refresh # Reindex all + rebuild + cleanup + vacuum
flo memory stats # Show statistics
Routing & Learning
flo hooks route --task "description" # Route task to optimal agent
flo hooks learn --pattern "..." --domain "." # Store a pattern
flo hooks patterns # List learned patterns
flo hooks consolidate # Promote/prune patterns
Gates
flo gate check-before-scan # Blocks Glob/Grep if memory not searched
flo gate check-before-agent # Blocks Agent tool if no TaskCreate
flo gate prompt-reminder # Context bracket tracking
flo gate session-reset # Reset gate state
Worktrees
flo worktree add feature/123-thing # Create a worktree and provision it (alias: flo wt)
flo worktree add feature/123 --json # {"path":…,"branch":…,"index":0,"provisioned":true}
flo worktree add feature/123 --from v2.1.0 # Branch off a specific ref
flo worktree add feature/123 --no-provision # Create it, skip copy/link/setup
flo worktree list # Every worktree and its provisioning state
flo worktree remove feature/123-thing # Refuses a dirty tree unless --force
Diagnostics
flo healer # Quick health check (environment, deps, config)
flo healer --fix # Auto-fix issues (memory DB, daemon, config, MCP, zombies)
flo diagnose # Full integration test (memory, swarm, hive, hooks, neural)
flo diagnose --suite memory # Run only memory tests
flo diagnose --json # JSON output for CI/automation
flo doctor is still accepted as an alias for flo healer — every flag and subcommand below works under either name.
flo healer — Health Check
flo healer runs 38 parallel health checks against your environment and reports pass/warn/fail for each:
Check
What it verifies
Version Freshness
Whether your installed MoFlo version matches the latest on npm (detects stale npx cache)
Node.js Version
Node.js >= 20 installed
npm Version
npm >= 9 installed
Claude Code CLI
claude command available
Git
Git installed
Git Repository
Project is inside a git repository
Config File
Valid moflo.yaml exists
Status Line
statusLine config wired in .claude/settings.json (auto-fixes when missing)
Daemon Status
Background daemon running (checks PID, cleans stale locks)
Memory Database
SQLite memory DB exists and is accessible
Embeddings
Vectors indexed in memory DB, HNSW index present
Embedding Hygiene
Indexer writes preserve existing embeddings instead of clobbering them
Test Directories
Test dirs from moflo.yaml exist on disk, reports auto-index status
MCP Servers
moflo MCP server configured in .mcp.json
Disk Space
Sufficient free disk space (warns at 80%, fails at 90%)
TypeScript
TypeScript compiler available
Semantic Quality
Semantic search returns relevant, varied results with acceptable similarity scores
Intelligence
SONA, ReasoningBank, PatternLearner, LoRA, EWC++, and RL subsystems are loaded
Spell Engine
Core spell modules, step commands, loaders, and index are present
Zombie Processes
No orphaned MoFlo node processes running
Subagent Health
Agent lifecycle (spawn → status → terminate) completes successfully
Spell Execution
End-to-end spell probe runs a real step and captures output
MCP Tool Invocation
MCP tool schemas are loaded and callable
MCP Spell Integration
Bridge between MCP tools and spell engine functions correctly
Hook Execution
Hook executor is functional and can fire hooks
Gate Health
All gate cases, hook bindings, and state file are intact
MofloDb Bridge
Memory DB adapter (node:sqlite + HNSW) is wired and routable
Sandbox Tier
Detects which sandbox backend is available (Docker / bwrap / sandbox-exec / none)
Auto-fix mode (flo healer --fix) attempts to repair each failing check automatically:
Issue
What --fix does
Missing memory database
Creates .swarm/ directory and initializes the SQLite DB
Embeddings not initialized
Initializes memory DB and runs embeddings init
Missing config file
Runs config init to generate defaults
Status line not wired
Adds statusLine config block to .claude/settings.json
Stale daemon lock
Removes stale .moflo/daemon.lock and restarts daemon
Malformed .mcp.json (e.g. unescaped Windows paths)
Backs up the broken file as .mcp.json.malformed-<ts>, regenerates a clean one, verifies it parses (#1126)
After auto-fixing, healer re-runs all checks and shows the updated results. Issues that can't be fixed automatically are listed with manual fix commands.
Additional flags:
flo healer --install # Auto-install missing Claude Code CLI
flo healer --kill-zombies # Find and kill orphaned MoFlo processes
flo healer -c memory # Check only a specific component
flo healer -c embeddings # Check only embeddings health
flo healer --verbose # Verbose output
/eldar — Consult the Eldar (project setup audit + wizard)
Where the Healer checks your moflo install, /eldar audits how Claude is set up to use the project — guidance, CLAUDE.md, memory namespaces, hook/MCP wiring, model routing, and stack-aware guidance gaps — then walks you through fixing whichever findings you pick. Use it when starting in a new project, when Claude feels lost or inefficient, or as a periodic health check.
/eldar # Read-only audit; categorized report + top-3 ranked recommendation
/eldar --fix # Audit, then interactive triage menu — pick which findings to address
The Eldar consult the Healer (they call flo healer --json as one of the audit checks) — they don't replace it. Categories audited include setup health, index freshness, version skew, model/token routing, CLAUDE.md size + reference integrity, guidance content + structure, memory health, hook/MCP wiring, settings sanity, spell + subagent inventory, stack → guidance cross-reference (detects tech from package.json/pyproject.toml/Cargo.toml/go.mod and flags every detected technology with no matching guidance doc — the highest-leverage finding for new adopters), and best-effort anti-pattern detection from history.
In --fix mode, each chosen finding drives the appropriate sub-flow: Healer for setup repair, the /guidance skill for guidance authoring (wizard, never autogen), a stack-aware scaffold for missing CLAUDE.md, flo init --upgrade for hook/MCP wiring. Every write is confirmed before it lands.
flo diagnose — Integration Tests
While healer checks your environment, diagnose exercises every subsystem end-to-end: memory CRUD, embedding generation, semantic search, swarm lifecycle, hive-mind consensus, task management, hooks, config, neural patterns, and init idempotency. All test data is cleaned up after each test — nothing is left behind.
GitHub Repository Setup
flo github setup # One-shot: generate CI + apply repo settings + branch protection
flo github setup --dry-run # Preview everything without making changes
flo github ci # Generate .github/workflows/ci.yml from project config
flo github ci --dry-run # Print CI config to stdout
flo github settings # Apply repo settings + branch protection via gh CLI
flo github settings --dry-run # Preview settings changes
flo github ci auto-detects your package manager (npm/pnpm/yarn/bun), TypeScript, and test directories from moflo.yaml and package.json, then generates a GitHub Actions CI pipeline with install, build, lint, type-check, and test steps.
flo github settings applies recommended defaults via gh CLI: delete-branch-on-merge, squash merge with PR title/body, auto-merge, linear history, and configurable branch protection (required reviews, dismiss stale reviews, block force pushes). Requires gh auth login.
System
flo init # Initialize project (one-time setup)
flo --version # Show version
What Ships Out of the Box
flo init wires up the following systems automatically. Here's what each one does, why it matters, and whether it's enabled by default.
Hooks (enabled OOTB)
Hooks are shell commands that Claude Code runs automatically at specific points in its lifecycle. MoFlo installs 26 hook bindings across 8 lifecycle events. You don't invoke these — they fire automatically.
Hook Event
What fires
What it does
Enabled OOTB
PreToolUse: Write/Edit
flo hooks pre-edit
Records which file is about to be edited, captures before-state for learning
Yes
PreToolUse: Glob/Grep
flo gate check-before-scan
Memory-first gate — blocks file exploration until memory is searched
Yes
PreToolUse: Read
flo gate check-before-read
Blocks reading guidance files directly until memory is searched
Yes
PreToolUse: Bash
flo gate check-dangerous-command
Safety check on shell commands
Yes
PreToolUse: Bash
flo gate check-before-pr
Validates PR readiness before gh pr create
Yes
PostToolUse: Write/Edit
flo hooks post-edit
Records edit outcome, optionally trains neural patterns
Yes
PostToolUse: Write/Edit
flo gate reset-edit-gates
Resets edit-related gate state after the write completes
Yes
PostToolUse: Agent
flo hooks post-task
Records task completion, feeds outcome into routing learner
Yes
PostToolUse: TaskCreate
flo gate record-task-created
Records that a task was registered (clears TaskCreate gate)
Yes
PostToolUse: Bash
flo gate check-bash-memory
Detects memory search commands in Bash (clears memory gate)
Yes
PostToolUse: Bash
flo gate record-test-run
Records test runs from Bash for the test-output gate
Yes
PostToolUse: Skill
flo gate record-skill-run
Records that a skill was invoked (clears skill-related gates)
Yes
PostToolUse: memory_search
flo gate record-memory-searched
Records that memory was searched (clears memory-first gate)
Your AI assistant accumulates project knowledge across sessions instead of starting from scratch each time
Yes
HNSW Vector Search
Hierarchical Navigable Small World index for fast nearest-neighbor search
Searches across thousands of stored entries return in milliseconds instead of scanning linearly
Yes
Guidance Indexing
Chunks markdown docs into overlapping segments, embeds each with MiniLM-L6-v2
Your project documentation becomes searchable by meaning ("how does auth work?") not just keywords
Yes
Code Map
Parses source files for exports, classes, functions, types
The AI can answer "where is X defined?" from the index instead of running Glob/Grep
Yes
Test Indexing
Maps test files to their source targets based on naming patterns
The AI can answer "what tests cover X?" and identify untested code
Yes
Gates
Hook-based enforcement of memory-first and task-registration patterns
Prevents the AI from wasting tokens on blind exploration and untracked agent spawns
Yes
Context Tracking
Interaction counter with bracket classification (FRESH/MODERATE/DEPLETED/CRITICAL)
Warns before context quality degrades, suggests when to checkpoint or start fresh
Yes
Semantic Routing
Matches task descriptions to agent types using vector similarity against 12 built-in patterns
Routes work to the right specialist (security-architect, tester, coder, etc.) automatically
Yes
Learned Routing
Records task outcomes (agent type + success/failure) and feeds them back into routing
Routing gets smarter over time — successful patterns are weighted higher in future recommendations
Yes
SONA Learning
Self-Optimizing Neural Architecture that learns from task trajectories
Adapts routing weights based on actual outcomes, not just keyword matching
Yes
MicroLoRA Adaptation
Rank-2 LoRA weight updates from successful patterns
Fine-grained model adaptation without full retraining
Yes
EWC++ Consolidation
Elastic Weight Consolidation that prevents catastrophic forgetting
New learning doesn't overwrite patterns from earlier sessions
Yes
Session Persistence
Stop hook exports session metrics; SessionStart hook restores prior state
Patterns learned on Monday are available on Friday
Yes
Auto-Meditate
Recognizes durable lessons in the live session and distills them into the learnings namespace in a background pass at the next session start
The high-signal lessons from each session are kept automatically — no need to remember to run /meditate
Yes
Status Line
Live dashboard showing git branch, session state, memory stats, MCP status
At-a-glance visibility into what MoFlo is doing
Yes
MCP Tool Server
80+ MCP tools for memory, hooks, coordination, spells, swarm, etc. (schemas deferred by default)
Lets Claude Code call MoFlo functionality directly
Yes (deferred)
Systems (available but off by default)
System
What It Does
How to Enable
Model Routing
Auto-selects haiku/sonnet/opus per task based on complexity analysis
model_routing.enabled: true in moflo.yaml
MCP Auto-Start
Starts MCP server automatically on session begin
mcp.auto_start: true in moflo.yaml
Tool Schema Eager Loading
Loads all MCP tool schemas (100+) at startup (instead of on-demand)
mcp.tool_defer: false in moflo.yaml
The Two-Layer Task System
MoFlo doesn't replace Claude Code's task system — it wraps it. Claude Code handles spawning agents and running code. MoFlo adds a coordination layer on top that handles memory, routing, and learning.
Here's how a typical task flows through both layers:
MoFlo routes — Before work starts, MoFlo analyzes the prompt and recommends an agent type and model tier via hook or MCP tool.
MoFlo gates — Before an agent can spawn, MoFlo verifies that memory was searched and a task was registered. This prevents blind exploration.
Claude Code executes — The actual agent runs through Claude Code's native task system. MoFlo doesn't manage the agent — Claude Code handles execution, output, and completion.
MoFlo learns — After the agent finishes, MoFlo records what worked (or didn't) in its memory database. Successful patterns feed into future routing.
The key insight: Claude Code handles execution, MoFlo handles knowledge. Claude Code is good at spawning agents and running code. MoFlo is good at remembering what happened, routing to the right agent, and ensuring prior knowledge is checked before exploring from scratch.
For complex work, MoFlo structures tasks into waves — a research wave discovers context, then an implementation wave acts on it — with dependencies tracked through both Claude Code's task system and MoFlo's coordination layer. The full integration pattern is documented in .claude/guidance/moflo-claude-swarm-cohesion.md.
The /flo skill ties both systems together for GitHub issues — driving the full process (research → enhance → implement → test → simplify → PR) with Claude Code's agents for execution and MoFlo's memory for continuity.
Intelligent Agent Routing
MoFlo ships with 12 built-in task patterns that map common work to the right agent type:
Pattern
Keywords
Primary Agent
security-task
auth, password, encryption, CVE
security-architect
testing-task
test, spec, coverage, e2e
tester
database-task
schema, migration, SQL, ORM
architect
feature-task
implement, add, create, build
architect → coder
bugfix-task
bug, fix, error, crash, debug
coder
api-task
endpoint, REST, route, handler
architect → coder
...
(12 patterns total)
When you route a task (flo hooks route --task "..." or via MCP), MoFlo runs semantic similarity against these patterns using HNSW vector search and returns a ranked recommendation with confidence scores.
The routing gets smarter over time. Every time a task completes successfully, MoFlo's post-task hook records the outcome — the full task description, which agent handled it, and whether it succeeded. These learned patterns are combined with the built-in seeds on every future route call. Because learned patterns contain rich task descriptions (not just short keywords), they discriminate better as they accumulate.
Routing outcomes persist across sessions. You can inspect them with flo hooks patterns or transfer them between projects with flo hooks transfer.
Memory & Knowledge Storage
MoFlo uses a SQLite database (via Node 22's built-in node:sqlite — no native deps) to store three types of knowledge:
Namespace
What's Stored
How It Gets There
guidance
Chunked project docs (.claude/guidance/, docs/) with 384-dim embeddings
flo-index on session start
code-map
Structural index of source files (exports, classes, functions)
flo-codemap on session start
tests
Test file → source target reverse mapping
flo-testmap on session start
patterns
Learned patterns from successful task outcomes
Post-task hooks after agent work
Semantic search uses cosine similarity on neural embeddings (MiniLM-L6-v2, 384 dimensions). When Claude searches memory, it gets the most relevant chunks ranked by semantic similarity — not keyword matching.
Session start indexing — Background indexers (guidance, code map, tests) run on every session start. They're incremental — unchanged files are skipped — and run in parallel so they don't block the session.
Cross-session persistence — Everything stored in the database survives across sessions. Patterns learned on Monday are available on Friday. The stop hook exports session metrics, and the session-restore hook loads prior state.
For Claude
When flo init runs, it appends a section to your CLAUDE.md that teaches Claude:
Always search memory before Glob/Grep/Read (enforced by gates)
Use mcp__moflo__memory_search for knowledge retrieval
Use /flo <issue> (or /fl) for issue execution
Store learnings after task completion
Full Configuration Reference
flo init generates a moflo.yaml at your project root. Here's the complete set of options:
project:
name: "my-project"
guidance:
directories: [.claude/guidance]
namespace: guidance
code_map:
directories: [src, packages]
extensions: [".ts", ".tsx"]
exclude: [node_modules, dist]
namespace: code-map
tests:
directories: [tests, __tests__]
patterns: ["*.test.*", "*.spec.*", "*.test-*"]
extensions: [".ts", ".tsx", ".js", ".jsx"]
exclude: [node_modules, coverage, dist]
namespace: tests
gates:
memory_first: true # Must search memory before file exploration
task_create_first: true # Must TaskCreate before Agent tool
task_status_gate: block # Open tasks block `gh pr create`; warn → report only; off → silent
context_tracking: true # Report measured context-window usage once per band crossing
# context_limit: 200k # Optional window size (200000 / 200k / 1m) to get a % instead of raw tokens
verify_before_done: true # On by default (#1294); /verify before `gh pr create` (unless --no-verify). false to disable
sdd:
default: false # true → every /flo run uses the spec→plan→implement→verify cycle unless --no-sdd
merge:
auto: false # true → every /flo run auto-merges its PR once preconditions pass, unless --no-merge
auto_index:
guidance: true # Auto-index docs on session start
code_map: true # Auto-index code on session start
tests: true # Auto-index test files on session start
auto_meditate:
enabled: true # Distill durable session lessons into the learnings namespace
mcp:
tool_defer: true # Defer MCP tool schemas (100+); loaded on demand via ToolSearch
auto_start: false # Auto-start MCP server on session begin
hooks:
pre_edit: true # Track file edits for learning
post_edit: true # Record edit outcomes
pre_task: true # Agent routing before task spawn
post_task: true # Record task results for learning
gate: true # Gate enforcement
route: true # Intelligent task routing
stop_hook: true # Session-end persistence
session_restore: true # Restore session state on start
models:
default: opus
research: sonnet
review: opus
test: sonnet
model_routing:
enabled: false # Set to true for automatic model selection
confidence_threshold: 0.85
cost_optimization: true
circuit_breaker: true
status_line:
enabled: true
branding: "MoFlo V4"
mode: compact # single-line, compact, or dashboard
show_dir: true # current directory name (compact/dashboard only)
show_git: true
show_session: true
show_swarm: true
show_mcp: true
Tool Deferral
By default, tool_defer is true. MoFlo exposes 100+ MCP tools — loading all their schemas at conversation start consumes significant context. With deferral enabled, only tool names are listed at startup (compact), and full schemas are fetched on demand via ToolSearch when actually needed. Hooks and CLI commands continue to work normally since they call the daemon directly, not through MCP tool schemas.
Set tool_defer: false if you want all tool schemas available immediately (useful for offline/air-gapped environments where ToolSearch may not work).
Model Routing
By default, MoFlo uses static model preferences — each agent role uses the model specified in models:. This is predictable and gives you full control.
Set model_routing.enabled: true to enable intelligent routing, which analyzes each task's complexity and auto-selects the cheapest capable model:
Complexity
Model
Example Tasks
Low
Haiku
Typos, renames, config changes, formatting
Medium
Sonnet
Implement features, write tests, fix bugs
High
Opus
Architecture, security audits, complex debugging
The router learns from outcomes — if a model fails a task, the circuit breaker penalizes it and escalates to a more capable model.
You can pin specific agents even when routing is enabled:
model_routing:
enabled: true
agent_overrides:
security-architect: opus # Never downgrade security work
researcher: sonnet # Pin research to sonnet
Architecture
9 bin entries shipped with npm: flo (main CLI), moflo, claude-flow (aliases), flo-setup, flo-codemap, flo-search, flo-embeddings, flo-index, flo-testmap
Project config system: moflo.yaml for per-project settings
One-stop init: flo init generates everything needed for OOTB operation
Background
MoFlo began as a fork of Claude Flow but is now an independent project. The two share some early roots and a few namespace conventions, but the codebases, runtime, and design priorities have fully diverged.
MoFlo narrows the focus to one scenario: day-to-day local coding with Claude Code on a single project. Instead of a broad, highly configurable orchestration framework, it bakes in sensible defaults, adds automatic indexing and memory gating at session start, and tunes the out-of-box experience so that npm install and flo init get you straight to coding. Over time it grew its own architecture — workspace collapse, in-tree fastembed runtime, node:sqlite + HNSW memory layer, spell engine, daemon-driven scheduling — and ships as a single npm package.
Contributing
Contributions are welcome — bug reports, documentation, tests, and code. See CONTRIBUTING.md for setup, conventions, and how to open a pull request. By participating, you agree to abide by our Code of Conduct.
License
MIT
M
Moflo
eric-cielo·19 ดาว·
MoFlo — an opinionated, local-first AI agent orchestration toolkit for Claude Code: semantic memory, learned routing, gates, and spells. No API keys, no cloud, works out of the box.
คัดลอกคอนฟิก stdio สำหรับ Cursor, Claude Desktop, Claude Code หรือ VS Code ได้ในหน้านี้ คอนฟิกนี้ใช้การติดตั้งของผู้เผยแพร่ ไม่ใช่พร็อกซีของ PlugMCP
Moflo ต้องใช้ API key ไหม?
ระเบียนนี้ไม่ได้ระบุตัวแปรสภาพแวดล้อมที่จำเป็น
Moflo ใช้ทรานสปอร์ตแบบใด?
Moflo ทำงานแบบโลคัลผ่าน stdio
Moflo ถูกตรวจสอบล่าสุดเมื่อใด?
ตรวจสอบล่าสุดเมื่อ 2026-10-28 MoFlo — an opinionated, local-first AI agent orchestration toolkit for Claude Code: semantic memory, learned routing, gates, and spells. No API keys, no cloud, works out of the box.