# MCP server

(Coming soon)

The Koo MCP server exposes your account to any MCP host — Claude Code, Codex, Cursor — as a set of `koo_*` tools. Through it, an agent can deploy a [service](/docs/apps/apps-and-services), follow the [deployment](/docs/apps/deployments-and-rollback) timeline to ready, roll back, tail [logs](/docs/observability/logs), read [metrics](/docs/observability/metrics), set variables, and check [status](/docs/observability/status) — without leaving the agent.

It authenticates with a single [API token](/docs/developers/api-tokens) (`kc_…`), scoping the agent to exactly one account and one role. The server never adds an account switch — the token is single-account by construction. Permissions are enforced by the API, not the agent: a `reader` token can watch a deploy but never start one.

> **Note:**
>
> The server is code-complete and verified in the Koo monorepo, but the `@koo-io/mcp-server` package is not published to npm yet, and the `koo mcp` CLI command ships with the next [CLI](/docs/developers/cli) release. Until both land, drive Koo from an agent over the plain REST [API](/docs/developers/api) with a `kc_…` token. The commands below are what will install it.

## Install

One command registers the server and its matching [skill](/docs/developers/ai-skills) in your host:

```text
koo mcp install claude-code   # or: codex, cursor
```

The install writes a host config block that launches the server with `koo mcp run` — no token is ever written into it. Your `kc_…` token is stored in your OS keyring and injected only at launch. Re-running the install is idempotent, and `koo mcp uninstall <host>` removes exactly what it added.

## Tools

Every tool returns structured content and acts on the single account the token is bound to. Reads are marked read-only; writes (`koo_deploy`, `koo_rollback`, `koo_set_variables`, and the creates) carry a destructive hint so your host can prompt before applying.

<!-- mcp-tools:begin -->

`koo_get_logs` and `koo_get_metrics` return the window actually queried after it is clamped to your plan's retention — the agent reports that resolved window and does not retry to widen it. Deploying from local source is not a tool: the skill routes that through `koo up` in the [CLI](/docs/developers/cli), then watches the result with these tools.

## Security

- 
- 
- 
-

## Until then

- 
-