overmind-core/overmind/overmind-agent

Overmind Agent

Start with list_projects and choose the intended accessible project. For an account connection, pass its project_id on every project tool and resource URI query; follow returned links. Project API keys retain their narrower access.

Turn the Agent surface into an evidence-based account of the project's capabilities, their contracts and the next improvement to make. The agent is the project; capabilities are its nodes, not separate agents.

Establish the map

Use the configured Overmind MCP connection. Read overmind://project/current?project_id=ID for project identity and repository provenance. A missing repository snapshot means the scanned revision is unknown; last_synced_at is not the scan time. Live traces can come from another revision.

Resolve a supplied capability through overmind://capabilities/{capability}. For project-wide orientation, get_instrumentation_plan without a capability returns registered placements and any required local action. Use returned identities; there is no list_capabilities tool. Placements and observed executions are evidence about the map, not proof of complete coverage.

Inspect the capability's status, task context, dataset, eval set, benchmark and live model. A leftover capability is an absent scan result, not a deletion request. MCP cannot create capability cards by scanning a repository.

Connect contracts to evidence

When prompts are supported, use investigate-capability with the resolved capability. Otherwise combine inspect_capability_health and query_failures as needed. Follow relevant trace resource links.

The capability resource lists the eval_matrix rows that score it and the Default set's eval_preload counts. Spans without a capability id indicate a mapping gap; low scoring coverage means spans arrive without evidence the matrix can judge. Missing scores do not demonstrate success.

Choose the next action from the evidence: refresh the repository map, instrument an uncovered entry point (Python codebases only; report any other language as not applicable), prepare data, repair evaluation coverage or investigate a measured failure. Do not automatically start those mutations during a map review.

Read capability state

A capability is the project-scoped product AI surface; use the word capability in MCP calls, with its eval matrix rows and scored traces underneath it. overmind://capabilities/{capability} returns safe metadata: name, slug, stored status, model, active dataset, active eval set, active model, selected benchmark and ready trained benchmark candidates. A reference may be a name, slug or id where the tool schema accepts it. get_instrumentation_plan with no capability returns placements for every current capability and inspect_capability_health reads one. The resource may report current, leftover or deleted; workflow tools resolve active current capabilities, and MCP does not reactivate or remove a capability.

Changes and handoffs

Repository discovery runs locally through the installed Overmind setup workflow (local repository setup) and overmind sync; MCP cannot inspect an arbitrary Git provider, create capability cards or edit files. If the project has no instrumentation plan or capability data, run local /overmind setup, preserve existing project and capability ids and configuration, run overmind sync, then re-read the capability resource and request get_instrumentation_plan. Never invent a capability id. If local tooling is unavailable, report that handoff and retain the current map.

benchmark_model selects the incumbent for future comparisons; active_model controls live serving. Preserve both during a map review. Use their respective mutation tools only for the user's requested selection.

• set_active_model checks a ready deployment, wakes it and verifies inference

before switching the capability alias. Poll the returned model_activation job with get_job; the old selection stays active until the job is complete. Read the capability resource for activation progress, previous selection and successful application traffic timestamps. Repeat the selection to retry a failed activation; omitting the deployment clears routing and cancels a pending switch. Fine-tune rollout uses get_model_swap_prompt only after a deployed fine-tune is ready, and applying that prompt stays a human action (Inference).

• set_benchmark_model selects a ready trained deployment for future incumbent

evaluations; read benchmark_candidates on the capability resource for the choices, and omit deployment to use the codebase incumbent. It is independent of set_active_model and does not change serving. The Console exposes the same selection on the capability's Models tab, not in training setup. New training jobs pin the selection; changing it does not retarget existing jobs.

Return the capability identities, observed coverage, provenance limitations and the next useful action. For a visual view, open the project's console_url; for one capability use capabilities/{id}, preserving its base and projectId.