TRAECNclaw MCP
# TraeCN MCP
TraeCN MCP connects your agent to the TraeCN desktop app on the same Mac.
Use 20 focused MCP tools to delegate work, follow results, change context and
settings, and resolve approvals. The local gateway manages the queue, task
recovery, and completion delivery.
```text
Agent host → TraeCN MCP (stdio) → local gateway → TraeCN (CDP)
```
## Quick start
Requires macOS, Node.js 22+, and a signed-in TraeCN desktop app. From this
checkout, install dependencies and start the gateway:
```sh
npm ci
npm run quickstart -- --launch-trae
```
Keep that terminal open while using the gateway. If TraeCN is already running
without remote debugging, follow the restart instructions printed by
Quickstart. It reuses your existing profile by default.
In another terminal, generate the host configuration:
```sh
npm run --silent setup:mcp
```
Add the generated `traecn` entry to your host's MCP configuration and reconnect.
See the [MCP server guide](docs/MCP-SERVER-GUIDE.md) for other config formats,
protocol compatibility, and tool details.
## Use it
Call `traecn_send_message` and keep the returned task ID. Use your host's
completion subscription or notification, or read the same task with
`traecn_get_task`. Select a workspace, conversation, model, or mode only when
the task requires a change.
## Operations and integrations
- **Web console** at [127.0.0.1:8788](http://127.0.0.1:8788/) for human control and task review.
- **CLI** for setup and diagnostics. Run `npm run doctor` or see the [deployment guide](docs/deployment.md) for background operation.
- **Agent Skill** for tool-use guidance. [Skill source](skills/traecn-mcp/SKILL.md)
- **OpenClaw and REST** use the same gateway. [OpenClaw setup](docs/OPENCLAW-MCP-SETUP.md) · [Runtime configuration](docs/configuration.md)
## Daily check-in
With a signed-in Trae CN 3.4.15 or 3.4.11 workspace connected, run `traecn-mcp check-in`
to read the account's daily status. Add `--claim` to claim the reward once, or
start the gateway with `TRAECN_AUTO_CHECK_IN=1` to check on startup and hourly.
For a LaunchAgent, set the same variable when installing or refreshing the
gateway service. Check-in uses Trae CN's Beijing business day and verifies
the result through the app. It uses the check-in action without the optional
draw. [Runtime configuration](docs/configuration.md)
## Distribution and development
Download the server tarball, MCPB bundle, or Agent Skill from
[GitHub Releases](https://github.com/Luckycat133/traecn-mcp/releases).
On macOS, Homebrew installs the same server tarball:
```sh
brew tap Luckycat133/traecn-mcp https://github.com/Luckycat133/traecn-mcp.git
brew install Luckycat133/traecn-mcp/traecnclaw
traecn-mcp quickstart --launch-trae
```
In another terminal, run `traecn-mcp setup-mcp` to generate your host entry.
Run `npm run verify` for local, non-live validation. See the
[compatibility matrix](docs/COMPATIBILITY.md),
[distribution runbook](docs/DISTRIBUTION.md), and
[channel ledger](docs/CHANNEL-LEDGER.md). For upgrades from 0.3.x, see the
[migration guide](docs/MIGRATING-TO-0.5.md).
TDQS
Scored across 20 tools
The tool set is largely cleanly partitioned: tasks, conversations, settings, models/modes, workspace, and interaction handling each have their own verbs. The only genuinely fuzzy pair is cancel_task vs stop_generation, but the description explicitly tells the agent to prefer cancel_task when a task ID exists, and the settings trio (toggle/option/text) is distinguished by control type. A few tools (list_settings vs list_setting_options vs list_setting_sections) are close cousins but the descriptions establish a clear required ordering.
Every tool uses the uniform traecn_<verb>_<noun> snake_case pattern (list_, select_, set_, create_, delete_, get_, cancel_, stop_, send_, open_, answer_, decide_). Verbs are used semantically consistently: list_ for read-only discovery, select_ for choosing from a listed set, set_ for applying a value. No camelCase or style mixing anywhere.
Twenty tools is on the heavy side, but the server spans several distinct sub-domains (task lifecycle, conversation lifecycle, settings discovery/mutation, model/mode selection, approvals, workspace), so most tools earn their place. It is borderline rather than bloated, and consolidation of the four settings tools would make it tighter.
Coverage is strong across most lifecycles: tasks (send/get/cancel), conversations (list/select/create/delete), settings (sections/list/options/set/select), models and mode, workspace opening, and interactive prompts (answer_question, decide_approval). Minor gaps remain — there is no list_tasks to enumerate tracked work, no conversation rename, and no workspace listing/current-workspace read — but agents can work around these using the returned taskId.