Exposes all CoReason platform operations as MCP tools. Enables upstream AI agents, IDEs, and orchestrators to control the platform natively. Delegates to src.core.services (same shared business logic as API/CLI/SDK).
Server exposes 24 tools with basic descriptions and schemas, but significant gaps undermine production readiness. Most tool descriptions are 1 - 2 sentences (avg ~80 chars), below the 194-char baseline for A+ tools. Parameter descriptions are minimal or absent (e.g., 'payload' in execute_agent lacks detail on expected structure). No output schemas documented, LLMs cannot predict response structure for chaining. Error handling is absent: tools return generic dicts with no recovery guidance. Security concerns: local_fs_writer/reader expose filesystem access without path validation or sandboxing. No tool annotations (readOnlyHint, destructiveHint) despite clear risk levels. Composition issues: execute_agent, resume_agent, and trigger_factory_build bundle multiple concerns. Parameter naming inconsistencies: 'payload' vs explicit fields elsewhere. No pagination support for list_projects despite potential scale. Overall, definitions are skeletal, suitable for internal use, not production agent integration.
Create a new project.
Delete a project by ID.
Trigger a LangGraph execution flow for a specified agent.
Export a project for air-gapped transfer.
Generate an MkDocs scaffold with config and markdown pages.
Get a specific agent's manifest and metadata.
Check the status of an enqueued agent execution.
No output schemas documented. LLMs cannot predict response structure, breaking tool chaining and forcing agents to guess field names. Every tool should document its return type (e.g., {status: string, project: {id, name, description}}).
Filesystem tools (local_fs_writer, local_fs_reader) lack path validation and sandboxing. No checks for directory traversal (../../../etc/passwd), symlink attacks, or permission boundaries. Agents can read/write arbitrary files on the host.
Parameter 'payload' in execute_agent, resume_agent, and trigger_factory_build is untyped object with no schema or description of expected structure. LLMs cannot construct valid payloads without trial-and-error.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 52 | <=2025-11-25 | v2 |
Retrieves the latest LangGraph state from the Postgres checkpointer.
Check the status of an async portability job (Push/Pull).
Fetch a single project by ID.
Get platform version info.
Check platform health — Postgres, Redis connectivity and version info.
Import a project from an air-gapped export bundle.
Injects a secret into the local dev Vault for agent impersonation.
List all agents in the factory with their metadata.
List all projects in the workspace.
Reads the content of a file from the local filesystem.
Writes the content string to the exact filepath on the local filesystem.
Pull a project from an OCI registry. Returns a job_id for async polling.
Push a project to an OCI registry. Returns a job_id for async polling.
Resumes a paused or failed agent execution by enqueuing a new task.
Rewind a session to a specific UUIDv7 checkpoint ID.
Tails the last N lines of the platform_worker logs.
Trigger a factory build for a new agent platform.
No error handling or recovery guidance. Tools return generic dicts (e.g., {error: 'Not found'}) with no actionable next steps. LLMs cannot distinguish retryable from fatal errors or know which tool to call next.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels. delete_project is DESTRUCTIVE but unmarked. Agents cannot reason about side effects or retry safety.