DevCouncil
Code intelligence and verification components for AI development.
Power your agent
การติดตั้งจากผู้เผยแพร่ ไม่ใช่เอนด์พอยต์ของ PlugMCP
URL แบบรีโมตหรือคำสั่งบนเครื่องของคุณ? การติดตั้งสองแบบต่างกันอย่างไร
รันบน PlugMCP
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.
เปิด"DevCouncil should not merely generate code. It should make AI-generated work prove that it satisfied the original intent."
Evidence, not model confidence, is the final authority.
DevCouncil is components and modules for AI-assisted software development. It ships independently installable binaries and libraries — code intelligence (devmap), task and lease state (dcstore), deterministic verification (dcverify), ignore-aware search (dcgrep), and a Go host (devcouncil / dev) — so a harness or an app can take only the pieces it needs and update them one at a time.
See repository input and verification boundaries for executable selection, filesystem protections, resource limits, and the difference between unavailable evidence and a passing check.
Manvi wraps these components into a coding-agent harness (turn loop, providers, policy, TUI, manvi serve). GitPulse uses Manvi for policy, workbench, and agent hosting, and DevCouncil components for code intelligence and related analysis. DevCouncil does not replace coding agents; it sits beside Claude Code, Codex, Cursor, and others as selectable modules.
DevCouncil's core runtime is a set of standalone, compiled native modules. Each binary is useful on its own, sits on PATH, and can be spawned or linked by a host that wants that one job:
flowchart TD
subgraph Consumers["AI Agents & Host Harnesses"]
Agent["Coding Agents (Cursor / Claude / Codex / Antigravity / Warp)"]
Manvi["Manvi (Agent Loop & Policy Harness)"]
GitPulse["GitPulse (Workbench & Git Host)"]
end
subgraph Orchestration["Go Host Layer"]
DevCouncil["devcouncil / dev (Go Host CLI & Stdio MCP)"]
end
subgraph NativeModules["Native Rust Engines"]
DevMap["devmap (Code Intelligence & AST Graph)"]
DCStore["dcstore (Task & Lease SQLite Store)"]
DCVerify["dcverify (Rigor, Stubs & Diff Verification)"]
DCGrep["dcgrep (Ignore-Aware Ripgrep Engine)"]
end
subgraph Storage["Persistent State"]
GraphDB[(".devmap/devmap.sqlite (Code Graph)")]
TaskDB[(".devcouncil/state.sqlite (Tasks & Leases)")]
end
Agent -->|"Code Navigation MCP"| DevMap
Agent -->|"Task & Policy MCP"| DevCouncil
Manvi -->|"Stdio JSON IPC"| DevCouncil
Manvi -->|"Stdio JSON IPC"| DevMap
GitPulse -->|"In-process / CLI"| DevMap
GitPulse -->|"Workbench state"| DCStore
DevCouncil -->|"Forward map/graph/ast"| DevMap
DevCouncil -->|"Task lease coordination"| DCStore
DevCouncil -->|"Rigor gate execution"| DCVerify
DevMap --> GraphDB
DCStore --> TaskDB
manvi serve) for policy, workbench, and agent hosting; selected DevCouncil crates and the devmap CLI for in-process code intelligence. It does not have to take every module.devmap alone, dcverify alone, the Go host MCP, or the full set. Updating one module does not require shipping the rest.[!NOTE] DevCouncil is built entirely as native compiled Go and Rust binaries; legacy Python orchestration has been retired (see docs/PHASE7_LONG_TAIL.md).
Six code-graph tools plus a ripgrep text baseline, measured on four repositories (595 to 4,335 files, mixing Rust, Go, TypeScript, Python and Swift) on one Apple M5 Pro. Every timing is a minimum or median from a reproducible run with the raw evidence committed beside it — not a marketing estimate. Head-to-head breakdown: DevMap vs GitNexus, CodeGraph, Graphify, Gortex, and codebase-memory-mcp. Full method, caveats, and raw output: benchmark report.
DevMap 0.2.2 was fastest on most of what was measured: cold indexing and unchanged refresh on all four repositories against the four tools timed as standalone CLIs (CodeGraph, codebase-memory-mcp, GitNexus, Graphify; Gortex is a resident daemon and is not comparable on those two stages), the lowest median in all six symbol-query cells, and 5/5 source-inspected caller pairs — matched only by codebase-memory-mcp. It did not win the one-file-changed refresh: CodeGraph did, on all four.
Detailed table on the 1,098-file DevCouncil corpus, where the correctness and query campaign also ran:
Where DevMap won. Fastest cold index on all four corpora (1.6–21× the time of the four CLI tools) and fastest unchanged refresh on all four (2.2–107×). Lowest median in all six query cells — definition lookups at 9.7 ms, ~10× faster than CodeGraph and ~82× faster than GitNexus. Lowest sampled peak memory on three of four corpora. All five source-inspected caller pairs, and a clean result on both index-staleness probes.
Where DevMap lost. CodeGraph re-indexes a single edited file faster on every corpus — DevMap takes 1.08× its time on the smallest and 3.61× on GitPulse. That is the one stage where a competitor is consistently ahead. DevMap's store is 1.6–2.0× CodeGraph's and 2.8–4.7× Graphify's, and Graphify used less memory than DevMap on the largest repository (1,458 MB vs 1,725 MB). Against v0.2.1, unchanged refresh regressed ~12% on the smallest corpus.
What these numbers are not. Four repositories, one machine, three-to-five repetitions, three symbols, and five inspected caller pairs. They do not establish general graph accuracy, persistent-MCP latency, or coding-agent task success. Tool output scopes differ, so equal latency is not equal analysis. DevMap itself reports 106,217 unexplained call-attribution sites on this corpus. The report states every limit explicitly and keeps the failures in — including GitNexus serving a deleted symbol through two ordinary refreshes until a forced rebuild.
Build just DevMap from source, then map a project:
git clone https://github.com/bharathvbcr/DevCouncil.git
cd DevCouncil
bash scripts/install.sh --only=devmap
export PATH="$HOME/.local/bin:$PATH"
cd /path/to/your/project
devmap build --manifest --guides
devmap paths --json
devmap status --json
devmap explore MyFunction --json
Git, Rust/Cargo and Go are required: scripts/install.sh builds the Go host
into a temporary directory and runs devcouncil install with your arguments,
and that command is the installer (--help lists presets and flags). Omit
--only=devmap for the Go host and full analysis suite. The host needs the
toolchain declared in go.mod, currently Go
1.26.9, and a sibling checkout of Gusset
at ../gusset, at the revision pinned in
.github/actions/setup-gusset/action.yml
(npm run ci:local and rust/verify.sh fail on any other revision). Without Go, cargo install --git https://github.com/bharathvbcr/DevCouncil.git --tag vX.Y.Z --locked devmap-cli, with vX.Y.Z a release tag (the default branch is whatever was merged last),
installs DevMap alone. The host installs only from a checkout: run from outside one, devcouncil install (which includes the host by default) exits 2 and names git clone https://github.com/bharathvbcr/DevCouncil.git plus devcouncil install --src <dir>; there is no go install route, and the npm package is a launcher for a host you built, not a source of one. On Windows use .\scripts\install.ps1 -Components devmap,
or omit -Components for the suite.
After a checkout install, the installer asks the apps that read DevCouncil's
schemas (GitPulse, Manvi, Jarvis) whether they carry this checkout's numbers.
It does this by running GitPulse's scripts/check-schema-parity.mjs with node,
from --gitpulse DIR, $DEVCOUNCIL_GITPULSE_ROOT or a sibling ../GitPulse.
Each stale or unverifiable reader is printed under FOLLOW-UP with its
remedy, and the install still succeeds. Pass --strict-consumers to fail it
instead, including when the check could not run. Without GitPulse or node, the
report says schema consumers: NOT CHECKED and why.
The npm package is a lightweight launcher for native binaries, not a bundled
runtime. Complete installation guide.
Use the paths returned by devmap paths: new repositories default to .devmap,
existing .devcouncil layouts remain supported, and explicit configuration can
override either. --guides requests managed workspace guides separately from
the map export. Check freshness, coverage gaps and result truncation before
using graph answers as evidence. If first-time guide creation leaves status
stale, run devmap build --manifest once more to index the new guides.
devmap integrate cursor --dry-run
devmap integrate cursor
devmap integrate cursor --check
Substitute claude, codex, antigravity, opencode or warp for cursor.
Integration can update user-level MCP settings as well as project assets.
Inspect the receipt and complete host reload/trust steps. For task tooling,
add the Go-host integration described in
coding CLI integration.
For an existing task in an initialized .devcouncil/state.sqlite:
devcouncil verify TASK-001 --mode enforce --json
flowchart LR
A["Acquire Task Lease<br/>(devcouncil_checkout_task)"] --> B["Implement Changes<br/>(Within Planned Scope)"]
B --> C["Verify Task<br/>(devcouncil verify / dcverify)"]
C --> D{"Rigor Gates<br/>& Scope Clean?"}
D -->|"Gaps Found"| E["Self-Repair<br/>(Typed next_actions)"]
E --> B
D -->|"All Gates Passed"| F["Release Task Lease<br/>(devcouncil_release_task)"]
Read gate mode, skipped reasons and coverage metadata. The Go host invokes
dcverify for rigor checks; changed-line coverage requires a profile supplied
through --coverage PATH. The default gate mode is off. Task leases and MCP
policy results coordinate participating clients; retired lifecycle hooks do
not intercept arbitrary shell commands or editor writes. The sandbox selector
does not implement Docker/Nix isolation. See the
task loop contract.
Start at the documentation index.
Licensed under Apache 2.0. The Python CLI, its release lineage and its historical workflow certifications are retired; current native source and installed binary identities must be checked separately.
tree-sitter/tree-sitter) (Max Brunsfeld & contributors, MIT) — Multi-language AST parsing in devmap.docs/benchmarks/ benchmark against GitNexus (Abhigyan Patwari), CodeGraph, Graphify, Gortex, and codebase-memory-mcp.Code intelligence and deterministic verification for AI coding agents: symbol-level code graph, blast radius, dead-code analysis, and MCP servers for Claude Code, Cursor, and Codex.
เครื่องมือ
ยังไม่มีรายการเครื่องมือ เซิร์ฟเวอร์แบบรีโมตจะได้รายการหลังการตรวจสอบสถานะครั้งแรก
FAQ
คัดลอกคอนฟิก stdio สำหรับ Cursor, Claude Desktop, Claude Code หรือ VS Code ได้ในหน้านี้ คอนฟิกนี้ใช้การติดตั้งของผู้เผยแพร่ ไม่ใช่พร็อกซีของ PlugMCP
ระเบียนนี้ไม่ได้ระบุตัวแปรสภาพแวดล้อมที่จำเป็น
DevCouncil ทำงานแบบโลคัลผ่าน stdio
ตรวจสอบล่าสุดเมื่อ 2026-11-05 Code intelligence and deterministic verification for AI coding agents: symbol-level code graph, blast radius, dead-code analysis, and MCP servers for Claude Code, Cursor, and Codex.