Skip to main content
Glama

Drift is a semantic version-control layer that wraps Git. Every commit becomes an Intent: the prompt that produced the change, the agent model behind it, the AST-level mutations, an optional checkpoint of the agent's cognitive state, and a cryptographic Ed25519 signature — all linked into an auditable, replayable graph.

Built for the AI era. When more than 80% of code is generated, text diffs are useless for review: they show what changed, never why. Drift rejects broken syntax before it enters history, answers "why does this function exist?" with the originating prompt, and lets a crashed agent resume from its last checkpoint.

License: MIT Node Tests

Quickstart

Give your agent Drift: Claude Code, Antigravity, Codex App, Codex CLI, Cursor, Factory Droid, Gemini CLI, GitHub Copilot CLI, Kimi Code, OpenCode, Pi.

Prefer no agent? Use the CLI, the GitHub App, the GitHub Action, or VS Code. Want the 5-minute "aha" first? Seed the demo repo and run drift blame.

Documentation: Quickstart (5-minute start) · API reference (CLI + MCP tools) · Architecture (how Drift works under the hood)

Verified live on Windows 11 (Node v24.18.0, 2026-08-06): fresh clone → first drift blame in ~8.1 s, 10/10 checks pass, no registry 404 — and the npm path (packed @drift/* chain installed into an empty dir) answers the MCP handshake with all six tools in ~1 s. See the full measured tables.

Related MCP server: gitwhy-mcp

How it works

It starts the moment you run drift init. Drift creates .drift/ — a SQLite DAG, a config, and a per-repo Ed25519 keypair — and from then on every commit becomes an intent.

When you (or your agent) run drift realize -p "<prompt>", Drift doesn't just commit. It parses the change semantically, rejects the commit if the syntax is broken (exit 2 — broken code never enters history), redacts secrets from your prompt, computes an AST delta (ADDED / MODIFIED / DELETED / MOVED / RENAMED), signs the intent, and stores it content-addressed in .drift/objects/ before committing with a Drift-Intent: trailer.

After that, drift blame can walk any line or function back to the prompt that created it, drift context hydrates the last intents for a file so an agent grounds itself before editing, and drift verify re-runs the recorded verification command. A crashed agent runs drift replay --checkout and resumes exactly where it left off.

And because these are MCP tools, your coding agent can use them directly — drift_realize instead of git commit.

Deeper reading: the full command reference lives in docs/api.md (CLI flags, exit codes, JSON schemas, MCP tool inputs), and docs/architecture.md explains the storage model, encryption at rest, the webhook app, and the security boundaries.

Installation

Installation differs by harness. If you use more than one, install Drift separately for each one. All harnesses expose the same six tools: drift_realize, drift_context, drift_replay, drift_blame, drift_verify, drift_log.

Status: the @drift/* npm packages are not published yet. Every section below leads with the clone path — it works right now from a checkout of this repository. The npx -y @drift/mcp / npx -y @drift/cli one-liners activate automatically once the packages land on npm; until then they return a 404, so use the clone command shown first.

Every command below is backed by a real manifest in this repository (.claude-plugin/plugin.json, .plugin/plugin.json, .cursor-plugin/plugin.json, .codex-plugin/plugin.json, gemini-extension.json, plugin.json, .factory-plugin/, package.jsonpi) or a ready-made config in examples/harness-configs/. Today, installation needs Node.js ≥ 24, npm and a clone of this repository (the MCP server runs straight from packages/drift-mcp/dist/index.js; no build step needed). Once the @drift/* packages are published, the same configs work via npx -y @drift/mcp with no clone.

Claude Code

Install as a plugin from the Drift marketplace (plugin-style, like Superpowers):

/plugin marketplace add lilcipherx/drift
/plugin install drift@drift

Or add the Drift MCP server directly (project scope) — from a clone:

claude mcp add drift --env DRIFT_REPO=/abs/path/to/your/repo -- node /path/to/drift/packages/drift-mcp/dist/index.js

Once the packages are published, the same command works via npx (no clone):

claude mcp add drift --env DRIFT_REPO=/abs/path/to/your/repo -- npx -y @drift/mcp

Or copy the ready-made config:

cp examples/harness-configs/claude-code/.mcp.json .mcp.json

Verify with claude mcp list — you should see drift with its six tools.

The marketplace manifest lives at .claude-plugin/marketplace.json in this repository (github-source lilcipherx/drift, strict plugin → .claude-plugin/plugin.json in the same repo).

Antigravity

Install Drift as a plugin from this repository:

agy plugin install https://github.com/lilcipherx/drift

Antigravity runs the plugin's session-start hook, so Drift is active from the first message. Reinstall with the same command to update.

Codex App

In the Codex app, open Settings → MCP servers and add:

  • Name: drift

  • Command: node (clone path) — or npx once the packages are published

  • Args: /path/to/drift/packages/drift-mcp/dist/index.js — or -y @drift/mcp after publication

  • Env: DRIFT_REPO=/abs/path/to/your/repo

Codex CLI

Add the Drift MCP server to ~/.codex/config.toml (from a clone):

[mcp_servers.drift]
command = "node"
args = ["/path/to/drift/packages/drift-mcp/dist/index.js"]
env = { DRIFT_REPO = "/abs/path/to/your/repo" }

Once published, the same server runs via npx (no clone):

[mcp_servers.drift]
command = "npx"
args = ["-y", "@drift/mcp"]
env = { DRIFT_REPO = "/abs/path/to/your/repo" }

Restart Codex, then ask for drift_blame / drift_context in any session.

Cursor

Copy the ready-made config and enable MCP servers in Cursor settings:

cp examples/harness-configs/cursor/mcp.json .cursor/mcp.json

Then ask for drift_blame in chat.

Factory Droid

  • Register the marketplace:

    droid plugin marketplace add https://github.com/lilcipherx/drift
  • Install the plugin:

    droid plugin install drift@drift

Gemini CLI

  • Install the extension:

    gemini extensions install https://github.com/lilcipherx/drift
  • Update later:

    gemini extensions update drift

GitHub Copilot CLI

Add the Drift MCP server (from a clone):

copilot mcp add drift -e DRIFT_REPO=/abs/path/to/your/repo -- node /path/to/drift/packages/drift-mcp/dist/index.js

Once the packages are published, the same command works via npx (no clone):

copilot mcp add drift -e DRIFT_REPO=/abs/path/to/your/repo -- npx -y @drift/mcp

Or copy the ready-made config to .github/mcp.json and restart Copilot.

Kimi Code

Drift is available in Kimi Code's plugin marketplace.

  • Open Kimi Code's plugin manager:

    /plugins
  • Go to Marketplace > Drift and install it.

  • Or install directly from this repository:

    /plugins install https://github.com/lilcipherx/drift
  • Detailed docs: docs/README.kimi.md

OpenCode

OpenCode uses its own plugin install; install Drift separately even if you already use it in another harness.

  • Tell OpenCode:

    Fetch and follow instructions from https://raw.githubusercontent.com/lilcipherx/drift/main/.opencode/INSTALL.md
  • Detailed docs: docs/README.opencode.md

Pi

Install Drift as a Pi package from this repository:

pi install git:github.com/lilcipherx/drift

For local development, run Pi with this checkout loaded as a temporary package:

pi -e /path/to/drift

VS Code

Add the Drift MCP server to .vscode/mcp.json (native VS Code MCP support):

cp examples/harness-configs/claude-code/.mcp.json .vscode/mcp.json

VS Code picks it up on window reload.

CLI

From a clone (works today):

git clone https://github.com/lilcipherx/drift.git && cd drift
npm install
node packages/drift-cli/dist/cli.js --help

Once the CLI is published to npm, npx works without cloning:

npx -y @drift/cli --help

On your own repository:

node /path/to/drift/packages/drift-cli/dist/cli.js init
# … edit a file …
node /path/to/drift/packages/drift-cli/dist/cli.js realize -p "Fix race condition in token refresh" --agent --model claude-3-5-sonnet
node /path/to/drift/packages/drift-cli/dist/cli.js log

GitHub App

Install @drift/app to get intent summaries on every pull request:

  • Reads Drift-Intent: trailers from PR commits, hydrates the intent objects from .drift/objects/ at the PR head, and posts a semantic summary comment — review the intent, not 2,000 lines of diff. Comments are idempotent: the app updates its own marker comment in place, so they never accumulate.

  • Runs as a webhook server: drift-app start (see packages/drift-app/app.yml for the app manifest, scripts/webhook-proxy.sh for local debugging).

GitHub Action

Check intent health in CI:

- uses: lilcipherx/drift@v0.3.0
  with:
    command: log     # or: doctor / verify <intent-id>

Demo

The 5-minute "aha" — a real Drift history, generated by the CLI itself:

bash scripts/seed-demo.sh
cd examples/demo-repo

node ../../packages/drift-cli/dist/cli.js log
node ../../packages/drift-cli/dist/cli.js blame src/auth.ts --function refreshToken

blame prints the prompt, model and a valid signature for the function:

src/auth.ts:12 (refreshToken)
  AGENT @ Drift Demo
  model:   claude-3-5-sonnet
  prompt:  Fix race condition in token refresh by de-duplicating in-flight refreshes
  intent:  did_2941b4547b4ed505a7c37190247768a7
  commit:  087c492f…  signature: valid

The Basic Workflow

  1. init — Creates .drift/ (SQLite DAG, config, Ed25519 keypair). Never rewrites history; deleting .drift/ leaves a fully functional git repo.

  2. realize — Commit with intent. Syntax gate (exit 2), secret redaction, AST delta, Ed25519 signature, Drift-Intent: trailer. This is git commit for the AI era.

  3. log — Timeline of intents: id, author (agent vs human), model, prompt.

  4. blame / contextblame --line|--function walks a symbol back to its originating prompt; context <file> hydrates the last N intents for grounding.

  5. verify — Re-runs the recorded verification command for an intent and checks the Ed25519 signature against the object file (never against DB rows).

  6. replay — Restore a checkpointed agent state; --checkout resets the worktree. Crash recovery for agents.

  7. doctor — DAG integrity, signature checks, orphan cleanup, encryption-key check when encryption is enabled.

The agent checks the intent before any task. Mandatory for anyone touching generated code.

What's Inside

Core

Package

What it does

@drift/cli

The drift CLI — init, realize, log, blame, context, verify, replay, doctor, export

@drift/core

Intent store (SQLite DAG), git wrapper, Ed25519 signatures, secret redaction, AES-256-GCM encryption at rest

@drift/ast

Semantic parser (TypeScript/JavaScript, Python) + AST deltas with a real syntax gate

Agent integration

Package

What it does

@drift/mcp

MCP server — six tools for Claude Code / Codex / Cline, delegates to the CLI

@drift/sdk

Typed SDK + Zod intent schemas

@drift/app

GitHub App — pull_request webhook that posts idempotent intent-summary comments

Review & CI

Package

What it does

@drift/action

GitHub Action (composite) — log / doctor / verify in CI

Philosophy

  • Semantics over text — diffs show what; intents show why. Always.

  • Broken code never enters history — the syntax gate is the front door.

  • Evidence over claims — every intent is signed; every claim is verifiable.

  • Security by default — secrets redacted, telemetry off, no network calls, optional AES-256-GCM encryption at rest (v0.2.0+).

  • Simplicity — zero native dependencies, strict git compatibility, no rewriting of history.

Security

  • Every intent is Ed25519-signed; verification uses the object-file canonical JSON, so signature checks never need the master key.

  • Prompts are regex-redacted for secrets (AWS, OpenAI, GitHub, Slack, JWT, PEM, …) before any storage.

  • Encryption at rest (v0.2.0): [encryption] enabled = true + DRIFT_MASTER_KEY encrypts prompt and agentState with AES-256-GCM (AAD-bound to the intent id). Note: the commit message keeps the plaintext prompt by design (PRD §9.1) — see SECURITY.md.

  • Keys are never committed (.drift/keys/ is gitignored) except throwaway demo keys.

Evaluation

The eval harness (PRD §22) drives the real CLI with mock file states — no LLM calls, no network — and records a baseline:

npm run eval           # run scenarios + compare against baseline (regression gate)
npm run eval:record    # re-record eval/baseline.json

Metrics gated at >5% regression (PRD §22.3): syntax-error rejection rate (must be 100%), blame accuracy, replay fidelity. Scenarios live in eval/scenarios/.

Contributing

See CONTRIBUTING.md. The general flow:

  1. Fork the repository.

  2. Create a branch for your work.

  3. Keep npm test green (110 tests: unit, temp-git-repo integration, MCP JSON-RPC e2e, GitHub App handler + live webhook-server E2E + client-abort

    • graceful-shutdown robustness).

  4. Keep the eval baseline green: npm run eval (PRD §22).

  5. Submit a PR using the template.

Design decisions are tracked in docs/adrs.md — note that the PRD originally chose Rust (ADR-003); this implementation ships TypeScript-first (ADR-006) for a zero-native-dependency MVS. The drift-ast parser interface is the drop-in point for a future tree-sitter implementation.

Updating

Once published, MCP servers launched via npx -y @drift/mcp pick up new versions automatically (npx always fetches the latest published release). Until then, servers launched from a clone (the commands shown in Installation) track this checkout — update with

git pull origin main
npm install

Releases are tagged on the releases page (v0.1.0, v0.2.0, v0.2.1, …).

Changelog: CHANGELOG.md.

License

MIT — see LICENSE. Security notes: SECURITY.md.

Community

A
license - permissive license
-
quality - not tested
A
maintenance

Maintenance

Maintainers
Response time
0dRelease cycle
6Releases (12mo)
Commit activity

Resources

Unclaimed servers have limited discoverability.

Looking for Admin?

If you are the server author, to access and configure the admin panel.

Related MCP Servers

  • A
    license
    B
    quality
    D
    maintenance
    Gives AI coding assistants persistent memory, safety controls, and project awareness by tracking coding sessions, protecting critical files from modifications, and managing approval workflows with automatic changelog generation.
    19
    20
    MIT
  • A
    license
    A
    quality
    A
    maintenance
    Change tracking for AI-era codebases. AI agents call it to log structured change events (entity + diff + reasoning) before the session ends, then query history with diff, blame, history, changeset, and search. Captures the intent that would otherwise evaporate.
    8
    19
    MIT

View all related MCP servers

Related MCP Connectors

  • Cross-agent artifact workspace with provenance across Claude Code, Codex, Cursor, LangGraph.

  • 33 tools that make AI write, implement, and verify intent against explicit, testable constraints.

  • Connect AI assistants to GitHub - manage repos, issues, PRs, and workflows through natural language.

View all MCP Connectors

Latest Blog Posts

MCP directory API

We provide all the information about MCP servers via our MCP API.

curl -X GET 'https://glama.ai/api/mcp/v1/servers/lilcipherx/drift'

If you have feedback or need assistance with the MCP directory API, please join our Discord server