TRAECNclaw MCP

Compose focused tools around user intent. Do not call preflight, wait, poll, event acknowledgement, recovery, cleanup, proof, process, or generic command operations from the agent loop; the gateway owns those mechanics.

Setup

1. Install the matching TRAECNclaw server from its source checkout or release server archive. The launcher checks the repository, an installed traecnclaw package, then traecnclaw-mcp on PATH; it never executes a JavaScript path selected by an environment variable.
2. Run npm run doctor in a source checkout. Then run npm run --silent setup:mcp from the checkout, or node scripts/setup-mcp.js from this Skill directory, to validate the server and print a resolved client entry. Use --output FILE to create a new mode-0600 file atomically; the command never overwrites an existing file. The JSON templates under assets remain manual fallbacks.
3. Start the gateway with npm run start:gateway, then reconnect the MCP client.
4. Use mock mode only for development; never cite mock output as live TraeCN evidence.

Choose only necessary context

• Call traecn_open_workspace only when the requested folder is not already open.
• Call traecn_list_models when the available choices matter; then call traecn_select_model once.
• Call traecn_select_mode only when the task requires a different Solo/IDE mode.
• Call traecn_list_conversations to read the current mode and stable conversation IDs.
• Use traecn_create_conversation, traecn_select_conversation, or traecn_delete_conversation for exactly one conversation operation.

These tools change TraeCN context. Never repeat workspace paths in the model message.

Send and track work

Call traecn_send_message with the exact message and, when concurrency makes the target ambiguous, an explicit conversationId. Clients without the Tasks extension receive the accepted taskId and status. When the host declares io.modelcontextprotocol/tasks on the request, the same call returns that durable gateway identity as resultType: "task".

Do not wait or busy-poll after sending. Legacy initialization-based clients receive compact task notifications that the stdio server acknowledges internally. Modern hosts can open subscriptions/listen for the returned task ID and receive a complete notifications/tasks snapshot; hosts without that extension call traecn_get_task only for an intentional later status/result read. If a modern task reports input_required, its elicitation and tasks/update response carry the same approval safeguards documented below. The first legacy connection for a client ID baselines retained history before notifications start. Reusing that stable ID later resumes only durably unacknowledged events; configure a distinct TRAECN_MCP_CLIENT_ID when the host-derived name is not unique. Use traecn_cancel_task for one known gateway task. Use traecn_stop_generation only when the user wants the generation currently visible in TraeCN stopped. First list conversations and pass the active conversationId; the gateway rejects a stale or different conversation. Because visible work may belong to the user or another agent, set acknowledgeUntrackedWork:true and include a short audit reason. Use traecn_cancel_task instead when the work has a gateway task ID.

For a completed task, omit detailLevel or pass "result" to retrieve only the final answer. Pass detailLevel: "trace" only when the detailed visible execution process is needed to diagnose or trace an unexpected outcome.

Inspect or change Trae settings

Treat settings as a second-level capability. Call traecn_list_setting_sections, then traecn_list_settings for only the needed section. For dropdowns, call traecn_list_setting_options instead of guessing. Use exactly one matching mutation tool:

• traecn_set_setting_toggle for a boolean switch.
• traecn_select_setting_option for a dropdown choice.
• traecn_set_setting_text for a text or numeric input.

These commands make the change through Trae's visible settings UI. The gateway serializes UI access and returns to chat automatically; do not navigate back manually or repeat reads after a successful write.

Resolve exceptional interactions

The gateway never approves shell commands automatically, including commands classified as read-only. Command text alone cannot establish workspace scope, exclude sensitive-file reads, or account for Git configuration and external hooks. Every command is returned to the Agent with its risk classification.

If a task reports an unresolved interaction:

• Use traecn_answer_question for a returned question and one of its offered answers.
• Use traecn_decide_approval only after checking the returned command and risk. Approval requires the exact expectedCommand, acknowledgeRisk:true, and a short audit reason; the gateway re-reads the visible card and rejects stale or changed commands. Denial needs no risk acknowledgement.

Conversation deletion is permanent. Delete only an inactive conversation and only when the current user request explicitly names or unambiguously identifies it for deletion. Re-read its listed ID and exact title, then pass the title with acknowledgePermanentDeletion:true; the gateway rejects title mismatches and the active conversation, records the attempt, and cannot restore a successful deletion. This uses the user's existing instruction and does not add a separate confirmation round trip.

Safety

• Keep the gateway on 127.0.0.1 unless remote access is secured.
• Never expose .env, tokens, authorization headers, or TraeCN profile data.
• Keep the default security audit log enabled; high-impact approvals, stops, and deletions are written to date-based JSONL files with sensitive fields redacted and a default 30-day retention.
• Treat mock results and non-terminal task states as incomplete.
• Do not recreate removed low-level tools through shell, generic commands, or GUI control.

Read references/mcp-surface.md only for exact schemas, protocol compatibility, notification behavior, and gateway ownership boundaries.