ContextGit

Every time you start a new session with an AI coding agent, it starts from zero. It doesn't know what was built yesterday, what decisions were made, or what's left to do. You spend the first 10 minutes of every session re-explaining your project. ContextGit fixes this — it gives your agents a persistent memory layer that survives across sessions, branches, and machines.

Install

npm install -g contextgit
cd your-project
contextgit init

Restart Claude Code (or any MCP client) after running init for the MCP server to take effect.

That's it. init registers the MCP server, updates CLAUDE.md with session instructions, and installs Claude Code hooks. Start a new session — the agent calls project_memory_load automatically and picks up exactly where the last session left off.

How it works

ContextGit stores structured context commits alongside your git history. Each commit captures what was done, what was decided, and what questions remain open.

Session 1: Agent builds auth module → saves context commit
Session 2: Agent loads snapshot → knows auth is done → starts on the next task

What the agent sees

When an agent calls project_memory_load, it gets a snapshot like this:

## Project State
Auth module complete. API routes tested. Database schema finalized.

## Current Branch: Context: main
Implementing payment integration. Stripe SDK configured.

## Recent Activity
- [2026-03-20T08:33:44Z] "Payment webhook handler done" by solo via claude-code
- [2026-03-20T07:15:22Z] "Stripe SDK integration" by solo via claude-code
- [2026-03-19T16:42:11Z] "Auth module complete" by solo via claude-code

## Open Threads
- [FREE] Need to add rate limiting to payment endpoints
- [CLAIMED by studio-mcp-agent] Build invoice PDF generation
- [FREE] Decide on webhook retry strategy

## Active Claims
- "Build invoice PDF generation" claimed by studio-mcp-agent (2h TTL)

The agent reads this and knows exactly where the project stands without you saying a word.

MCP tools

These tools are exposed to the agent via MCP. The agent calls them as part of its normal workflow.

CLI commands

contextgit init              # Initialize contextgit in current project
contextgit init --hooks      # Initialize with git hooks
contextgit commit "message"  # Write a context commit
contextgit status            # Show current project state
contextgit search "query"    # Search past commits
contextgit claim "task"      # Claim a task
contextgit unclaim <id>      # Release a claim
contextgit branch <name>     # Create a context branch
contextgit merge <id>        # Merge a context branch
contextgit serve             # Start the REST API server

Architecture

Monorepo with strict dependency ordering:

packages/
  core/     → Types, snapshot formatter, embedding service, context engine
  store/    → Storage interface, LocalStore (SQLite), RemoteStore, SupabaseStore
  mcp/      → MCP server (stdio transport, launched by Claude Code)
  cli/      → CLI commands (oclif)
  api/      → REST API (Express)

Local storage: SQLite via better-sqlite3. One DB per project at ~/.contextgit/projects/<id>.db.

Remote storage: Optional Supabase (Postgres + pgvector) for cross-machine sync and team use.

Embeddings: Local all-MiniLM-L6-v2 (384 dimensions, no API key needed) for semantic search.

Configuration

Project config lives at .contextgit/config.json:

{
  "project": "my-project",
  "projectId": "unique-id",
  "store": "local",
  "agentRole": "solo",
  "workflowType": "interactive",
  "autoSnapshot": false,
  "snapshotInterval": 10,
  "embeddingModel": "local"
}

Current status

This is an early-stage tool. It works for solo developers using Claude Code across one or more machines. The core memory layer is solid — context commits, branches, threads, search, and snapshots all work.

Known limitations

• No automatic context merge when git branches are merged (manual project_memory_merge required)
• No integration with external boards (Linear, Jira, GitHub Issues)
• Semantic search requires the embedding model to download on first use (~23MB)

License

MIT