MCP server that delegates Claude Code agent work to local models (llama.cpp / vLLM)
Server has 9 tools with complete input schemas and descriptions. Naming follows verb_noun convention (list_models, run_agent, wait_job, etc.). Descriptions are detailed (100-300 chars) and explain WHEN to use each tool. However, output schemas are not formally documented, responses are described in prose but lack structured JSON Schema definitions. Parameter descriptions are present but some lack constraints (e.g., isolation accepts 'none' or 'worktree' but no enum). Error handling guidance is minimal; tools do not explain recovery paths. Security considerations (credentials, permissions) are not addressed in tool definitions.
Cancel a running job (kills the headless Claude session).
Tail of a job's human-readable transcript (assistant text, tool calls, results).
Current status, last activity, files touched, and result (if finished) of a job.
List jobs from this server session (newest last).
List configured endpoints with live health and what each is serving right now. Per endpoint: `up`, `models` (served ids), `context` (real per-request window per served id), and slot/queue occupancy where the backend exposes it. Anything listed as served is usable immediately via `run_agent`; if nothing suitable is up, ask the user to start a model.
One-shot generation with a local model (no tools, no agent loop).
Output schemas not formally documented. Tool descriptions mention returned fields (e.g., 'up', 'models', 'context' for list_models) but no JSON Schema definition exists. LLMs cannot reliably extract structured data or plan downstream calls without explicit output schemas.
Parameter constraints not formalized as enums or patterns. 'isolation' accepts 'none' or 'worktree' but is defined as a free-form string. 'permission_mode' lists options in description but not as enum. LLMs may hallucinate invalid values.
Error handling lacks recovery guidance. Tools do not document what errors are retryable, user-fixable, or fatal. ModelUnavailable exception is caught but no guidance on next steps (e.g., 'ask user to start a model'). Agents cannot self-correct.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | <=2025-11-25 | v2 |
Add or update an inference endpoint (persisted to models.local.yaml) and probe it. `base_url` is the server root that exposes `/v1/messages`, e.g. `http://127.0.0.1:8080`. `backend` is one of llama.cpp | vllm | ollama | other (informational).
Delegate a task to an agent on a local model. Returns a job id immediately. The agent is a headless Claude Code session (Read/Edit/Write/Glob/Grep/Bash, project CLAUDE.md, same working tree) whose API calls go to the local endpoint. Write the task like a brief for a capable but junior engineer: what to do, where, how to verify, and what to report.
Block up to `timeout_s` seconds for a job to finish; returns its status/result either way.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Tools are marked with Risk field (READ_ONLY, WRITE, REVERSIBLE) in metadata but not exposed via MCP tool annotations. Agents cannot infer safety properties from the protocol.
Pagination not documented for list_jobs. Tool accepts include_finished boolean but no limit, offset, or cursor parameters. Large job lists could exhaust context. No guidance on result limits.