@nones/kaiten-mcp

MCP (Model Context Protocol) server for Kaiten — lets Claude Code, Claude Desktop, Cursor and other MCP clients manage your Kaiten boards: cards, child cards, comments, tags, checklists, files and card-scoped git branches. Ships as a single kaiten binary: kaiten mcp starts the MCP server, every other command is a regular CLI; a typed SDK is exported for programmatic use.

Migrating from v3? The separate kaiten-mcp binary was removed in v4.0.0 — append mcp to the command in your MCP config and launch it with npx -y @nones/kaiten-mcp mcp.

MCP clients (recommended setup)

You need two values:

| Variable | Where to get it | | --- | --- | | KAITEN_API_URL | https://<kaiten_url>/api/latest | | KAITEN_API_TOKEN | Kaiten → Profile settings → API tokens → Create token |

If they already live in the project's .env, initialize Kaiten once from that project:

npx -y @nones/kaiten-mcp init

init reads the existing runtime sources (process.env, .kaiten.env, .env, or an existing global config), then saves only the API URL and token to ~/.kaiten/config with file mode 0600. It does not print the token. Keep .env out of version control.

The MCP entry then needs no credential paths, shell wrapper, or project working directory:

{
  "mcpServers": {
    "kaiten": {
      "command": "npx",
      "args": ["-y", "@nones/kaiten-mcp", "mcp"]
    }
  }
}

The same entry shape works in Codex Desktop, Claude Code/Claude Desktop and Cursor; place it in that client's MCP configuration file.

Codex and PWD: Codex may start the MCP process with a working directory different from your project. Credentials saved by init do not depend on PWD. Set an explicit absolute cwd only if you intentionally keep optional project-specific defaults in ./.kaiten.env or ./.env; do not rely on the inherited PWD for those files.

Codex Desktop exposes these tools in task chats. After updating the package or MCP configuration, restart Codex Desktop and begin a new chat so it refreshes tools/list.

Run kaiten_health after connecting to verify only safe metadata (sources, readiness and normalized API origin).

Secret safety

Tokens and config-file contents are never returned by kaiten_health, MCP errors, or diagnostics. Prefer kaiten init or an MCP client secret store; do not put tokens in URLs or commit them to the repository.

Alternative: MCP-managed environment

If you do not want a user-level ~/.kaiten/config, pass credentials through the MCP client instead. For Claude Code:

claude mcp add kaiten \
  -e KAITEN_API_URL=https://<kaiten_url>/api/latest \
  -e KAITEN_API_TOKEN=your_token \
  -- npx -y @nones/kaiten-mcp mcp

For Claude Desktop, Cursor, and other clients, add them to the MCP config (claude_desktop_config.json, .cursor/mcp.json, …):

{
  "mcpServers": {
    "kaiten": {
      "command": "npx",
      "args": ["-y", "@nones/kaiten-mcp", "mcp"],
      "env": {
        "KAITEN_API_URL": "https://<kaiten_url>/api/latest",
        "KAITEN_API_TOKEN": "your_token"
      }
    }
  }
}

That's it — ask your assistant to list your Kaiten spaces to verify the connection.

Note for AI agents: if kaiten tools fail with a "Missing required Kaiten configuration" error, the server is installed but not configured. Show the error to the user — it explains exactly which variables (KAITEN_API_URL, KAITEN_API_TOKEN) to set and where.

Destructive operations

Every tool backed by an HTTP DELETE to the Kaiten API — kaiten_delete_card, kaiten_delete_checklist, kaiten_delete_checklist_item, kaiten_delete_comment, kaiten_remove_tag, kaiten_detach_from_parent, kaiten_unblock_card, kaiten_remove_external_link — is disabled by default. Calling one returns an actionable error instead of deleting anything.

Deleting Kaiten entities is irreversible. To enable DELETE tools, a human must add KAITEN_API_DELETE_ALLOW=true to the server's environment on purpose — for example in the env block of the MCP config:

{
  "mcpServers": {
    "kaiten": {
      "command": "npx",
      "args": ["-y", "@nones/kaiten-mcp", "mcp"],
      "env": {
        "KAITEN_API_URL": "https://<kaiten_url>/api/latest",
        "KAITEN_API_TOKEN": "your_token",
        "KAITEN_API_DELETE_ALLOW": "true"
      }
    }
  }
}

Or in ~/.kaiten/config / ./.kaiten.env / ./.env (see Configuration).

Note for AI agents: if a DELETE tool fails with a "DELETE operations are disabled for safety" error, do not set KAITEN_API_DELETE_ALLOW yourself — relay the error to the user verbatim and let them decide whether to enable it.

Agent guardrails (v3.3.0+)

Optional environment variables that confine what an AI agent can see and do through this server. All are off by default — with none set, behavior is unchanged. Configured guardrails compose as an intersection: an operation must pass every check.

| Variable | Effect | | --- | --- | | KAITEN_READ_ONLY | true blocks every create/update/delete at the HTTP layer — the agent can only read | | KAITEN_ALLOWED_SPACE_IDS / KAITEN_ALLOWED_BOARD_IDS | allowlists: listings are filtered, operations outside are rejected | | KAITEN_ALLOWED_COLUMN_IDS / KAITEN_ALLOWED_LANE_IDS | same, per column / lane | | KAITEN_PROTECTED_COLUMN_IDS | agent cannot move/create cards into these columns (e.g. Done/Deploy), reading is allowed | | KAITEN_ENFORCE_SCOPE | true turns KAITEN_DEFAULT_SPACE_ID/KAITEN_DEFAULT_BOARD_ID into hard boundaries — one-variable "single board" confinement | | KAITEN_DISABLED_TOOLS | comma-separated tool names to hide from the agent entirely | | KAITEN_CARDS_LIMIT | cap card-list responses at N cards (truncation is reported explicitly) | | KAITEN_MINIMAL_DEFAULT | true makes card listings token-minimal by default | | KAITEN_DEFAULT_COLUMN_ID / KAITEN_DEFAULT_LANE_ID | where kaiten_create_card puts cards when the agent omits column/lane | | KAITEN_HTTP_TIMEOUT_MS / KAITEN_HTTP_RETRIES | per-request timeout; retries for reads on transient failures |

Example — a read-only agent locked to one board:

"env": {
  "KAITEN_API_URL": "https://<kaiten_url>/api/latest",
  "KAITEN_API_TOKEN": "your_token",
  "KAITEN_DEFAULT_SPACE_ID": "12345",
  "KAITEN_DEFAULT_BOARD_ID": "67890",
  "KAITEN_ENFORCE_SCOPE": "true",
  "KAITEN_READ_ONLY": "true"
}

Guardrails constrain the agent using this server — they are not a substitute for the API token's own permissions. Full semantics, composition rules and enforcement tiers: Configuration → Guardrails.

Note for AI agents: never set, change or remove guardrail variables yourself. If an operation fails with an error naming one of these variables, relay the error to the user verbatim and let them decide.

Documentation

• Configuration — all variables, config files, precedence, allowlists
• MCP tools — all 58 tools and their token-saving flags
• CLI — the kaiten command reference
• SDK — programmatic usage from Node/Bun
• Architecture — layers, Kaiten Core, design decisions
• Development — building, testing, releasing

Requirements

Node.js >= 18. No other runtime dependencies beyond the MCP SDK.

License

MIT