AnchorKit

Homepage: https://anchorkit.dev

AI-native coding CLI with a persistent, project-scoped knowledge vault.

AnchorKit wraps the AI agents you already use (Claude Code, Cursor, or direct API access to Anthropic / OpenAI / Gemini) and adds:

• a Markdown vault under .anchorkit/vault/ that survives across sessions — decisions, incidents, module maps, runbooks
• slash commands for spec-driven work, code review, security review, vault search, and more
• an MCP server you can plug into Claude Code so the same vault is visible to both tools

Install

npm install -g @qtv/anchorkit

Requires Node ≥ 20. The native SQLite binding ships as a prebuilt binary for Linux x64/arm64, macOS x64/arm64, and Windows x64; on other platforms npm install will compile it (Python + a C++ toolchain required).

You also need at least one AI backend on your PATH or configured via an API key:

| Backend | How | |---|---| | Claude Code CLI | claude login | | Cursor CLI | install + sign in | | Anthropic API | export ANTHROPIC_API_KEY=... | | OpenAI API | export OPENAI_API_KEY=... | | Gemini API | export GEMINI_API_KEY=... |

Pick one inside the REPL with /connect.

Quick start

cd your-project
anchorkit onboard      # first time: scaffolds .anchorkit/vault/
anchorkit              # opens the REPL

Inside the REPL:

• /help — list every command
• /connect — choose the AI backend
• /vault-map — index the codebase into .anchorkit/vault/
• /vault-query <query> — semantic search across the vault
• /commit — generate a conventional-commit message from the staged diff
• /review / /security-review — code review and OWASP review of the diff
• /exit — quit

Resume a session

anchorkit --resume                # resume the most recent session
anchorkit --resume <session-id>   # resume a specific saved session

Sessions live under .anchorkit/state.json per project.

Use AnchorKit as an MCP server inside Claude Code

anchorkit internal-mcp-server boots a stdio-based MCP server that exposes the vault tools (VaultQuery, VaultMap, VaultWrite, TodoWrite, plan-mode helpers) to any MCP client. The typical wiring is via Claude Code's .claude/mcp.json:

{
  "mcpServers": {
    "anchorkit": {
      "command": "anchorkit",
      "args": ["internal-mcp-server"]
    }
  }
}

After that, Claude Code can read/write your vault and reuse the same context AnchorKit's REPL builds up.

What's in the vault

anchorkit onboard lays out an Obsidian-compatible Zettelkasten under .anchorkit/vault/:

.anchorkit/vault/
├── CLAUDE.md          ← project rules and conventions (auto-injected)
├── PROJECT.md         ← project vision
├── STATE.md           ← persistent state across sessions
├── knowledge/         ← concepts, patterns, module docs
├── decisions/         ← ADRs
├── incidents/         ← debugging sagas, root causes
├── flows/             ← workflows, pipelines, runbooks
├── maps/              ← Maps of Content (index notes)
├── archive/           ← superseded knowledge
└── _sessions/         ← session lifecycle notes

Everything is plain Markdown, so the vault is portable and reviewable in git.

Beta status

This is a 1.0.0-beta release. The API surface (slash commands, vault schema, MCP tools) is stable; cloud sync features are wired but disabled by default. Set ANCHORKIT_SYNC_DISABLE=1 to silence sync errors when you have no credentials.

Privacy & telemetry

When you are logged in to an AnchorKit account, the REPL sends a minimal heartbeat to the backend so we can count active beta users (Prometheus gauge anchorkit_users_online). The payload is exactly:

• machineId — a random UUID stored at ~/.anchorkit/machine-id (file mode 0600), generated once per machine and never tied to your hardware
• clientKind — the literal string "cli"
• platform — process.platform (e.g. linux, darwin, win32)
• cliVersion — the version printed by anchorkit --version

Nothing else is collected. No prompts, commands, file contents, file paths, vault notes, repo names, git remotes, environment values, hostnames, or usernames leave the machine through this channel.

The heartbeat:

• only fires when you are logged in and the REPL is running — never in anchorkit onboard, --version, or any one-shot subcommand;
• pings on a backend-advised cadence (default 60 s, clamp ≥ 30 s);
• fails silently — network errors back off exponentially and never block or degrade any feature.

Opt-out

Set the environment variable before launching the REPL:

export ANCHORKIT_DISABLE_HEARTBEAT=1

When opt-out is on at boot:

• no heartbeat is sent under any circumstance;
• the ~/.anchorkit/machine-id file is not created if it does not already exist (LGPD minimisation — no identifier is provisioned for a user who declined telemetry);
• a pre-existing machine-id is preserved untouched so re-enabling telemetry later does not inflate the install count;
• everything else in the CLI continues to work exactly the same.

The env var is read once at boot. Restart the REPL to flip the flag.

Legal basis: legítimo interesse (LGPD art. 7, IX + art. 10) for technical usage metrics. Details: https://github.com/QTVentures/anchorkit-ai/blob/main/backend/PRIVACY.md.

Architecture

Contributing code? Read ARCHITECTURE.md first: it describes the layers under src/, the import rules, and the rules every change must respect (for example, the vault is written only through the harness).

License

Apache-2.0. See LICENSE.

Links

• Repository: https://github.com/QTVentures/anchorkit-ai
• Issues: https://github.com/QTVentures/anchorkit-ai/issues