agent-context-graph

agent-context-graph is a 100% local MCP server for AI coding agents working inside a codebase. It builds a compact knowledge graph of files and symbols, enforces declared edit scope before writes, and records append-only reasoning logs so future agents and maintainers can understand why changes happened.

No frontend is required. The human-facing interface is this README plus the command-line tool.

flowchart LR
  Project["Target project"] --> MCP["agent-context-graph MCP server<br/>stdio only"]
  Codex["Codex CLI"] --> MCP
  Claude["Claude Code"] --> MCP
  Cursor["Cursor"] --> MCP
  MCP --> Graph[".agent-context-graph/graph<br/>nodes.jsonl + edges.jsonl"]
  MCP --> Cache["cache.sqlite<br/>local rebuildable cache"]
  MCP --> Scope["scope-lock engine"]
  MCP --> Logs["logs/*.jsonl<br/>reasoning history"]

Privacy & Local-First Guarantee

Everything runs on the user's machine. There is no SaaS service, telemetry, analytics, update check, crash reporting, GitHub API call, or external runtime API call. The only intended network activity is dependency installation through npm install.

The test suite includes test/noNetwork.test.ts, which scans compiled JavaScript for forbidden runtime network APIs.

Related MCP server: codeweave-mcp

Requirements

• Node.js 20 or newer.

• npm.

• Git, if you are cloning from GitHub.

• Windows, macOS, or Linux.

This project uses better-sqlite3, a native Node package. Most users should receive a prebuilt binary during npm install. If npm has to build it locally, your machine may need normal C/C++ build tools for your platform.

Quick Start From GitHub

Clone the repository:

git clone https://github.com/Raviraj2024/Agent-Context-Graph.git
cd agent-context-graph

Install dependencies and build:

npm install
npm run build
npm test

Run the CLI from the cloned repo:

node dist/bin/agent-context-graph.js status

For easier local testing, link the command globally on your machine:

npm link

After linking, this should work from any directory:

agent-context-graph status

To remove the local global link later:

npm unlink -g agent-context-graph

Use It In A Target Project

A target project is the codebase you want an AI coding agent to work on.

Go to that project:

cd /path/to/your/target-project

Initialize the graph:

agent-context-graph init

Check status:

agent-context-graph status

Connect your MCP client:

agent-context-graph connect codex
agent-context-graph connect claude-code
agent-context-graph connect cursor

You only need to run the connect command for the client you actually use. The command merges an MCP server entry into the project-local config file for that client.

Start the server manually for a smoke test:

agent-context-graph serve

In normal use, your MCP client starts agent-context-graph serve for you.

If You Do Not Want To Use npm link

You can run the built CLI directly from the cloned repository.

From a target project:

node /absolute/path/to/agent-context-graph/dist/bin/agent-context-graph.js init
node /absolute/path/to/agent-context-graph/dist/bin/agent-context-graph.js status
node /absolute/path/to/agent-context-graph/dist/bin/agent-context-graph.js connect codex

On Windows PowerShell, an example path looks like:

node C:\Users\you\Projects\agent-context-graph\dist\bin\agent-context-graph.js init

Client Setup

Antigravity

Antigravity is an agent-first IDE, so it is a good place to test this project. The exact MCP settings screen or config path may vary by Antigravity version, but the server values are the same anywhere MCP stdio servers are supported.

First prepare this tool:

git clone <agent-context-graph-repo-url>
cd agent-context-graph
npm install
npm run build
npm link

Then go to the project you want Antigravity to work on:

cd /path/to/your/target-project
agent-context-graph init

In Antigravity's MCP server settings, add a stdio server with these values:

{
  "name": "agent-context-graph",
  "command": "agent-context-graph",
  "args": ["serve"],
  "cwd": "/absolute/path/to/your/target-project"
}

If you do not use npm link, use the direct Node command instead:

{
  "name": "agent-context-graph",
  "command": "node",
  "args": ["/absolute/path/to/agent-context-graph/dist/bin/agent-context-graph.js", "serve"],
  "cwd": "/absolute/path/to/your/target-project"
}

On Windows, use escaped backslashes or forward slashes in JSON paths:

{
  "name": "agent-context-graph",
  "command": "node",
  "args": ["C:/Users/you/Projects/agent-context-graph/dist/bin/agent-context-graph.js", "serve"],
  "cwd": "C:/Users/you/Projects/my-target-project"
}

Restart Antigravity after adding the MCP server. Then use a prompt like:

Use the agent-context-graph MCP server before editing.
First call get_project_overview, init_or_refresh_graph, query relevant best practices,
declare the task scope, check scope before edits, and record every change.

Now implement: <your task>

For example:

Use the agent-context-graph MCP server before editing.
First call get_project_overview, init_or_refresh_graph, query relevant best practices,
declare the task scope, check scope before edits, and record every change.

Now add password reset support to the auth module.

If Antigravity shows the MCP server as connected, the agent should be able to call the graph, scope-lock, and logging tools automatically during development.

Codex

From your target project:

agent-context-graph connect codex

This writes or updates:

.codex/config.toml

It adds an MCP server named agent-context-graph with serve as the command.

Claude Code

From your target project:

agent-context-graph connect claude-code

This writes or updates:

.claude/mcp.json

Cursor

From your target project:

agent-context-graph connect cursor

This writes or updates:

.cursor/mcp.json

Expected Agent Workflow

MCP clients read the server instructions at session start. The intended workflow is:

1. Call get_project_overview.

2. Call init_or_refresh_graph.

3. Call query_best_practices for the relevant domain before implementing.

4. Call declare_task_scope before editing files.

5. Call check_scope before every file write.

6. Stop and ask the user if a proposed change returns needs_approval or hard_stop.

7. Call record_change immediately after every create, modify, or delete.

The server does not show its own approval prompt. It returns structured decisions so the host agent can ask the human.

CLI Reference

agent-context-graph init
agent-context-graph connect <codex|claude-code|cursor>
agent-context-graph serve
agent-context-graph status
agent-context-graph reset

• init: scans the current project, writes defaults, builds the graph snapshot, loads standards, and populates the local SQLite cache.

• connect <codex|claude-code|cursor>: merges an MCP server entry into the selected project-local client config.

• serve: starts the stdio MCP server.

• status: prints snapshot/cache counts and cache metadata.

• reset: deletes only cache.sqlite, then rebuilds cache state from the snapshot/current files.

MCP Tools

• get_project_overview: root node, detected stack, entry points, and node/edge counts.

• init_or_refresh_graph: builds or refreshes the graph and reports changed files.

• get_node_context: returns one node plus immediate incoming/outgoing edges.

• get_blast_radius: traverses calls/imports/inherits with static and inferred results separated.

• get_definitive_path: returns an ordered minimal change path for rename, signature, delete, or logic changes.

• query_best_practices: returns bundled standard nodes for a requested domain.

• declare_task_scope: creates the scope an agent intends to edit.

• check_scope: classifies proposed writes as auto-allowed, needs approval, or hard stop.

• record_change: appends a scrubbed reasoning log entry after a file mutation.

• get_node_history: returns compact prior reasoning for a node or path.

Project Data Directory

When used inside a target project, this tool creates .agent-context-graph/:

• config.json: project-local configuration.

• graph/nodes.jsonl: canonical committed graph nodes.

• graph/edges.jsonl: canonical committed graph edges.

• logs/<timestamp>__<session-id>.jsonl: committed append-only reasoning logs.

• change-index.json: compact per-path history.

• cache.sqlite: local rebuildable cache, do not commit.

• .lock: temporary cache write lock, do not commit.

Target projects should commit:

.agent-context-graph/config.json
.agent-context-graph/graph/nodes.jsonl
.agent-context-graph/graph/edges.jsonl
.agent-context-graph/logs/*.jsonl
.agent-context-graph/change-index.json

Target projects should ignore:

.agent-context-graph/cache.sqlite
.agent-context-graph/cache.sqlite-*
.agent-context-graph/.lock

Supported Source Files

The initial adapters index:

• TypeScript: .ts, .tsx

• JavaScript: .js, .jsx, .mjs, .cjs

• Python: .py

The graph stores signatures, docstrings, line ranges, tags, and hashes. It never stores full file contents or full function bodies.

The default scanner skips node_modules, .git, common build output folders, binary files, .agent-context-graph, and files larger than 1 MB.

Best-Practices Knowledge Base

The server includes local markdown standards under src/knowledgeBase/content/ for:

• backend architecture

• API design

• auth and RBAC

• security

• testing strategy

• error handling and observability

• data modeling

These are loaded into the graph as standard nodes and returned by query_best_practices.

Verifying A Clone

After cloning and building this repository, run:

npm run build
npm test

Expected result:

Test Files  7 passed
Tests       10 passed

You can also test init against a temporary project:

cd /path/to/some/project
agent-context-graph init
agent-context-graph status

You should see .agent-context-graph/config.json, .agent-context-graph/graph/nodes.jsonl, and .agent-context-graph/graph/edges.jsonl.

Troubleshooting

agent-context-graph command not found

Run this from the cloned repo:

npm link

Or use the direct Node command:

node /absolute/path/to/agent-context-graph/dist/bin/agent-context-graph.js status

npm install fails on better-sqlite3

Use Node 20 or newer. If your Node version is very new and a prebuilt binary is not available, npm may try to compile locally. Install your platform's native build tools, then rerun:

npm install

On Windows, this may require Visual Studio Build Tools with C++ support. On macOS, this may require Xcode Command Line Tools. On Linux, this may require Python, make, and a C++ compiler.

status shows no graph nodes

Run:

agent-context-graph init

Make sure you are inside the target project, not inside an unrelated parent directory.

MCP client does not show the server

Run the relevant connect command again from the target project:

agent-context-graph connect codex

Then restart the MCP client so it reloads local config.

Cache looks stale or broken

Run:

agent-context-graph reset

This removes only the rebuildable SQLite cache. It does not delete graph JSONL snapshots or logs.

Design Decisions

• Incremental refresh reports changed files, then rebuilds the graph snapshot in one pass. This keeps v1 deterministic and avoids stale cross-file import edges while preserving the external refresh contract.

• Parser adapters are isolated behind LanguageAdapter. The current adapters use conservative syntax scanning and include web-tree-sitter as the required parser dependency so WASM grammar-backed extraction can be expanded without changing graph or server code.

• Connector commands write project-local config files (.codex/config.toml, .claude/mcp.json, .cursor/mcp.json) and merge only the agent-context-graph entry.

• The graph schema has no body/content field. Only signatures, docstrings, line ranges, tags, and hashes are persisted.

Adding a Language Adapter

1. Add a new adapter in src/parsers/ that implements LanguageAdapter.

2. Return file/symbol nodes with stable qualified names and contains edges from file to symbols.

3. Keep source bodies out of nodes and edges.

4. Register the adapter in src/parsers/index.ts.

5. Add fixture coverage in test/fixtures/ and parser tests.

Development

npm install
npm run build
npm test

The test suite covers path safety, parser extraction, graph refresh, scope-lock precedence, secret scrubbing, lock handling, and the no-runtime-network static check.

License

MIT