에이전트 강화
게시자가 제공하는 설치 방식입니다. PlugMCP 엔드포인트가 아닙니다.
원격 URL인가요, 내 컴퓨터에서 실행하는 명령인가요? 두 설치 방식의 차이
PlugMCP에서 실행
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
실행게시자 제공
Run your CLI coding agents (Claude Code, Codex, Gemini CLI, OpenCode…), local models with Ollama, and HTTP providers such as LiteLLM, NVIDIA and OpenRouter from one self-hosted dashboard, with shared MCPs, secrets and repeatable workflows.
Smaller prompts, more code where code is enough: fewer hallucinations, lower token bill, eco-design by default.
Self-hosted control plane for AI coding agents (Claude Code, Codex, Gemini, Ollama). Shared MCPs, reusable workflows, AI audit that makes any repo AI-friendly. Engineering, not prompting : smaller prompts, fewer hallucinations, lower token bill.
Status: 0.15.1 (current release). Functional but pre-1.0. Breaking changes happen between minor versions; patch versions are safe. License: AGPL-3.0. Using Kronn locally to build your own product is fine; the copyleft only kicks in if you distribute a modified Kronn to others. See License notes.
0.15.1 makes the AI-doc audit trustworthy, lets you test prompt and skill versions on real runs before shipping them, and gives agents local media.
Earlier releases are described in the CHANGELOG.
If you've ever:
Kronn is a self-hosted control plane for AI coding agents: Claude Code, Codex, OpenCode, Vibe, Gemini CLI, Kiro, GitHub Copilot CLI, Ollama, and LiteLLM. Backend in Rust, frontend in React, secrets in an encrypted vault on your machine.
The interface is available in English, French, Spanish and Simplified Chinese. Its language is independent from the output language assigned to new agent discussions.
Already happy with one CLI? Kronn pays off the moment you add a second agent or want to replay a prompt across N files. It's the layer above your single assistant.
Most "AI-powered" workflows rely on one mega-prompt asking the agent to do 12 things at once. Expensive in tokens. Flaky in output. Opaque when it breaks. Non-deterministic by nature.
Kronn flips that:
ApiCall to fetch, JsonData for fixtures, Exec for shell, Notify for webhooks, BatchApiCall to fan out N HTTP calls in parallel. Zero tokens, deterministic, debuggable.BatchQuickPrompt to fan it out in parallel discussions.Prerequisite: at least one local CLI agent, Ollama for fully-local models, or an existing LiteLLM proxy. Kronn drives the runtime you already use; it does not ship its own LLM. The setup wizard auto-detects what's on your $PATH with an npx runtime fallback for npm-published agents.
Download the installer for your OS from Releases. No Docker, no extra runtime. Per-OS steps in docs/install.md.
git clone --branch 0.15.1 --depth 1 https://github.com/DocRoms/Kronn.git # latest stable release
cd Kronn
./kronn start # guided setup & launch (Docker)
macOS: use
./kronn start-devinstead (native). Docker on macOS cannot run your host agents (Claude, Codex, …) or read the Keychain —start-devruns backend + UI natively in one command. (make installalso lands on the guided setup.)
start-dev is the contributor/development mode with backend hot-reload. It
requires Rust, Node ≥ 24, pnpm and
Watchexec:
# macOS
brew install node watchexec
# Linux / WSL (after Rust is installed)
cargo install --locked watchexec-cli
The command checks these tools before starting and names every missing
executable with an install hint. Watchexec is development-only: the desktop
app, Docker deployment and a bare make run-backend do not require it.
Requires Docker + Docker Compose. On Windows, WSL2 (Docker Engine inside WSL works, Docker Desktop optional).
git clone --branch 0.15.1 --depth 1 https://github.com/DocRoms/Kronn.git # latest stable release
cd Kronn
./kronn start
# → http://localhost:3140
In both modes the setup wizard scans your repos and walks you through API keys.
Day 1: first 5 minutes
.mcp.json files and adopts MCPs automatically.{{variables}}. One-shot or fan-out (Batch) over a list of items.When you want orchestration (multi-step, conditional, scheduled, gated), graduate to a Workflow.
Eight flows that scratch real itches. Each one is a primary use case, pick whichever maps to your day.
Save Analyse {{ticket}} and draft a PR as a Quick Prompt. Paste 30 Jira keys (or pull them from a tracker step). Kronn fans out one discussion per item, in parallel, each with its own git worktree if you want isolation. Land them all behind a single human-approval Gate so nothing merges without you.
Concrete run (real numbers from a 30-ticket batch on Claude Sonnet, ~4k input + 6k output tokens per ticket, with prompt caching on): ~$3 of API spend, ~3 min wall time, 0 manual copy-paste. Same batch on Ollama with
qwen3.8:27brunning locally: $0, a few minutes of wall time depending on your hardware.
Before you wire a prompt into a workflow, click Compare on any Quick Prompt. Every target uses Kronn's shared agent + model picker: compare Claude, Codex, Gemini, Vibe or Ollama, or two tiers of the same agent. The rendered prompt runs in parallel and rich Markdown answers appear side by side in a dedicated workspace. Each column keeps a durable discussion you can open to continue the work.
If they all converge on the same answer → your prompt is solid, ship it into a Batch or a Workflow step. If they diverge → your prompt is ambiguous, fix it before it costs you 30× on a batch run.
Compare-agents = the QA gate of prompt engineering. Cheap to run once, saves a fortune if it catches a bad prompt before it enters production.
Create ideas or tasks in the global Planning workspace, rank them inside Critical / High / Normal / Low bands, and link the same task to several projects or discussions without duplicating it. Each discussion gets a compact plan panel with its primary objective, active/later work, subtasks and Definition of Done checklists. Humans and agents update the same durable state; every change keeps its actor and agents can fetch only the task or delta they need through MCP instead of rereading the whole conversation.
Project task lists deep-link straight to the full Planning detail, so the backlog, project view and discussion plan always lead back to one source of truth.
Run Llama 3, Gemma, Qwen, Codestral on your machine via Ollama as a first-class agent. Same Discussions / Quick Prompts / Workflows surface, Kronn just routes the calls to http://localhost:11434/v1/ instead of the cloud. Pick your default model right from Settings. Zero token bill, zero code leaving your laptop.
Already route models through LiteLLM? Declare the endpoint and optional key, test the connection, then map the models reported by your proxy to Kronn's Economy / Default / Reasoning tiers. Ollama and LiteLLM can both call Kronn's native tools from a discussion.
A real Auto-Dev workflow in Kronn looks like:
ApiCall fetches the Jira ticket → 0 tokensAgent (Claude, small focused prompt validated via compare-agents) drafts the implementation → N tokens, the only step where reasoning mattersExec runs cargo test → 0 tokensAgent (small focused prompt) reviews the diff against the tests → N tokens, boundedNotify posts the result to Slack → 0 tokens5 steps, 3 deterministic. AI fires only on the 2 reasoning steps, each with a small bounded prompt. Result: predictable cost, predictable output, debuggable when it breaks. See The Kronn way.
You build this via the wizard UI (drag-drop step types, autocomplete on agent/API/QP references) and can export/import Kronn's versioned JSON bundle. The workflow model deliberately shares Symphony's four workspace-hook names, but Kronn does not currently import WORKFLOW.md files.
Workflow engine supports 12 step types total: Agent, ApiCall, BatchApiCall, BatchQuickPrompt, JsonData, CollectApiData, TransformData, PublishPageData, Notify, Gate (human approval), Exec (allowlist-gated process execution), and SubWorkflow for composing reusable workflows. The three data-pipeline steps collect, reshape and publish typed JSON without an agent. Full reference in docs/architecture/overview.md and Live Pages architecture.
A Live Page is a shared, versioned HTML report—not a monitoring backend. Its sandboxed document can render Markdown, tables and charts while a workflow updates only its persisted JSON datasets. Pages can also start standalone or with mock data, then be linked to workflows and discussions later. A button on a Page can launch a Quick Prompt, Quick API, Quick Exec or workflow for each row of its data, shows that row's live state, and opens a card telling what the run produced.
Most "ask an AI about my repo" flows restart cold every conversation. Kronn flips that: the first time you onboard a project, an audit pass reads your source + configs and writes a structured docs/ tree that lives in the repo.
docs/AGENTS.md: stack, common tasks, source-of-truth files, code-placement rulesdocs/repo-map.md: folder tree + key entrypointsdocs/coding-rules.md: linters, formatters, conventions actually observeddocs/testing-quality.md: coverage, untested components, smoke commandsdocs/architecture/overview.md: patterns, data flow, legacy migrations tabledocs/operations/debug-operations.md: commands + troubleshootingdocs/operations/mcp-servers.md: per-MCP capabilities discovered via introspectiondocs/glossary.md: domain glossary built from your actual codedocs/tech-debt/*.md + docs/inconsistencies-tech-debt.md: one detail file per finding, severity-tagged, all linked from a single index table.Unknowns get marked or. A validate phase walks you through them interactively, project status moving NoTemplate → TemplateInstalled → Bootstrapped → Audited → Validated.
The whole tree is injected as project context into every Discussion, Quick Prompt and Workflow on that project, so your agent doesn't restart from scratch every chat.
Drift detection is granular: each section tracks its source files (e.g. coding-rules.md watches package.json + tsconfig.json + rustfmt.toml). When those drift, the section is flagged as stale and you can re-audit just that step. No re-running the whole audit when one config changed.
What 0.8.2 hardens (from real-world misses)
StrictHostKeyChecking, .env* tracking/deploy hygiene without reading secret values, web a11y/CSP). Each item produces an explicit "verified present / verified absent / TD" line so you trust the audit didn't just look the other way.audit_history YAML block on each TD detail file tracks every pass. A reconciliation report (_reconciliation-<date>.md) classifies dropped TDs as Fixed / Stale / Missed / Uncertain so nothing silently vanishes between audits.Verified in source (the agent opened the file and confirmed) vs Inferred (pattern-matched only). Phase 3 validation skips Verified ones to save your time.Security, Docker, Performance, Accessibility, Database, ApiDesign, CodeQuality) in one launch (16 steps total), each opening with a relevance gate ("not applicable → one line and move on"). The focused kinds — plus Drift, Rgaa and a Custom escape hatch — remain launchable standalone. Each one writes to its own index file so it doesn't clobber the Full audit's TDs.audit_runs with duration, severity counts (Critical × 12 + High × 4 + Medium × 1.5 + Low × 0.3 = score 0–100) and a recommendations_json list. When ≥3 findings cluster in one dimension (e.g. 5 Docker TDs), the dashboard surfaces "Run a focused Docker audit" inline.LICENSE, SECURITY.md, CONTRIBUTING.md, CODE_OF_CONDUCT.md, issue/PR templates. Private projects skip this block entirely.This is what makes Kronn a knowledge-persistence layer with strict baselines, not just a prompt launcher.
The audit isn't a doc that sits in a folder. 0.8.2 wires it into action.
When the AI confirms KRONN:VALIDATION_COMPLETE on the validation discussion, the page surfaces two CTAs:
ticket-to-pr preset already applied AND configured for your project:
repo_url (github.com → mcp-github, gitlab.com → mcp-gitlab, …){owner}/{repo} path params filled in from the parsed remotefetch_issue) switched from JSON fixture → real ApiCall against your trackerfetch_issue → analyze → plan_gate → implement → run_tests → review → create_pr → ready_gate → notify_done).The 9-step preset uses Kronn's désagentification discipline: only analyze, implement and review burn agent tokens. The rest is mechanical (ApiCall, Exec, Gate, Notify).
The killer pattern this unlocks: create a GitHub issue from your phone → AutoPilot's Tracker trigger picks it up → the workflow drafts the PR while you're away from the keyboard, pausing at plan_gate for your approval and at ready_gate before the PR is merged. Real autonomy with two human checkpoints.
The run_tests step is typically npm test, cargo test, composer test, pytest, … — all of which need vendored deps that don't exist in a fresh git worktree (no node_modules, no vendor, no target). Each Exec step now has an optional Setup phase that fires immediately before the main command:
☑ This step runs in a git worktree — install dependencies first
Setup command: composer install --no-interaction --prefer-dist
(auto-detected from composer.json — edit if needed)
Main command: bin/phpunit -c phpunit.xml.dist
Same allowlist + same timeout. Setup fails → main is NOT executed and you see the install logs. Setup succeeds → main runs in the prepared worktree. Common patterns are pre-listed (composer, npm ci, pnpm, yarn, poetry, pip) — pick one, override if needed.
For dockerized projects, the docker-in-docker volume mismatch is solved at the infra layer (self-mount + host-path translation), so docker compose run --rm svc <cmd> works from a worktree without any extra config.
AI agents confidently assert file paths that don't exist, functions they invented, versions that aren't yours. When one of those lies gets persisted into docs/AGENTS.md, every future agent inherits it as "established truth". 0.8.7 ships a two-tier guard that runs on every agent reply.
Three modes, defaulting to warn, picked in Settings → Sourcing & Anti-hallucination (or config.toml) :
off — no directive, no checking. The pre-0.8.7 behaviour.warn (default) — each agent receives a sourcing directive : "cite a file:line, URL, or user-confirmed for every non-trivial claim; if you can't, say so plainly and verify before asserting." Kronn then mechanically inspects each [src: …] citation the agent emits — path-jailed to the project root (no FS escape), checks the file exists and the line range is in bounds. A non-blocking per-message pill surfaces problems : red for fabricated citations (the cited file/line does not exist, or the source type is training-data = model's prior knowledge, auto-rejected), amber for confident claims left without any anchor. Clicking the pill opens a detail panel listing the bad citations.enforce (0.8.8) — warn, plus refuses writes to AGENTS.md curated="ai" sections whose citations don't resolve.Honest by design : verified means the citation exists, not that the claim is true (catching "real but irrelevant" citations is the job of a future LLM-judge layer). The pill is non-blocking — it's a signal you decide to act on, not a hard gate. Language-agnostic + compression-proof + ungameable : you can't fabricate a file or line number that actually exists in your repo.
The full convention is an open spec (backend/docs/conventions/agents-md-format-v1.md, served on /api/conventions/agents-md-format-v1 and linked from the Settings section) — anyone can emit Kronn-compatible documentation with or without Kronn.
docs/AGENTS.md) plus a structured audit tree (docs/glossary.md, docs/repo-map.md, docs/architecture/overview.md and friends) covering legacy hotspots, test gaps, coding conventions and per-MCP capabilities. Each section has drift detection: re-audit only what's stale.[+ Inviter] button, each agent talks via MCP tools (disc_append, disc_wait_for_peer), no human messenger needed. Validated live on a 3-agent roleplay session.{{variables}} and conditional sections. One-shot, fan-out, or chained.Three independent layers injected into the agent's system prompt:
17 default profiles, 25 default skills, custom ones via Markdown + YAML in ~/.config/kronn/.
MCP / API / hybrid plugins: configure once with encrypted secrets, link to projects N:N. MCPs sync to the right config file per agent (.mcp.json for Claude Code, ~/.codex/config.toml for Codex, .gemini/settings.json for Gemini, etc.). API plugins inject endpoints + auth into the system prompt so the agent calls them via Bash curl, no MCP server needed. OAuth2 client-credentials handled transparently.
Any REST API works. 40+ built-in plugins (GitHub, GitLab, Jira, Stripe, Slack, Chartbeat, Adobe Analytics, AWS CloudWatch…) cover the common cases, and a Custom API option lets you describe any other vendor in a freeform form: name, base URL, description, optional docs link, and the credential fields the agent needs. An AI helper bubble pre-fills the form from a curl example or a docs URL in one click.
Fits if you run multiple AI CLIs, want to decompose AI work into bounded workflows (PR review, audit, generate-from-spec, batch over tickets), where mechanical steps are deterministic code and AI only fires on reasoning. Bonus: self-hosted with secrets in your own vault.
Doesn't fit for pure RAG pipelines, in-process Python orchestration (LangGraph is the better path), or workloads needing managed multi-tenancy. Kronn is a mixed-step workflow engine (CLI agents + deterministic code) on local files, not a Python LLM framework, not a multi-tenant SaaS.
Vs other tools:
Self-hosted means we owe you precision on where your data goes.
~/.config/kronn/config.toml (file mode 0600, dir 0700 on Unix; per-user ACLs on Windows). Filesystem-permission protection; no OS keychain or Argon2id passphrase yet. Agent prompts go to the CLI vendor you chose (Anthropic / OpenAI / Google / Mistral / nobody if you pick Ollama). Kronn itself never phones home. Full data-flow diagram in docs/operations/auth-and-tls.md.$HOME read-only and ~/.<agent> config dirs read-write: exactly what's needed to drive agents, nothing more. Detailed mounts list in docs/install.md.sh -c, argv literals only. Auth bypass + TLS migration plan in docs/operations/auth-and-tls.md.In practice, for the 99% of users:
Full text in LICENSE. If your legal team wants more nuance, dual-licensing discussions are open.
Self-hosted control plane for AI coding agents (Claude Code, Codex, Gemini, Ollama). Shared MCPs, reusable workflows, AI audit that makes any repo AI-friendly. Engineering, not prompting : smaller pro
개요
Self-hosted control plane for AI coding agents (Claude Code, Codex, Gemini, Ollama). Shared MCPs, reusable workflows, AI audit that makes any repo AI-friendly. Engineering, not prompting : smaller prompts, fewer hallucinations, lower token bill.
도구
아직 도구 목록이 없습니다. 원격 서버는 첫 상태 확인 후에 도구 목록이 생깁니다.
FAQ
이 페이지에서 Cursor, Claude Desktop, Claude Code 또는 VS Code용 stdio 설정을 복사합니다. PlugMCP 프록시가 아니라 게시자의 설치 방식을 사용합니다.
이 레코드에는 필수 환경 변수가 등록되어 있지 않습니다.
Kronn은(는) stdio를 통해 로컬에서 실행됩니다.
마지막 확인: 2026-11-05. Self-hosted control plane for AI coding agents (Claude Code, Codex, Gemini, Ollama). Shared MCPs, reusable workflows, AI audit that makes any repo AI-friendly. Engineering, not prompting : smaller prompts, fewer hallucinations, lower token bill.
