Gives Claude Code real tools for a supervised run (status, the morning report, checkpoints, validators, the quota-window check) instead of only slash commands. Shells out to the local `agentrelay` binary, scoped to a workspace directory.
AgentRelay MCP server has well-structured tool definitions with clear, domain-specific descriptions (avg 180 chars, within baseline 194). All 7 tools have explicit input schemas with typed parameters and descriptions. However, output schemas are entirely undocumented, LLMs cannot predict return structure. Parameter descriptions are good but lack format constraints (e.g., 'minutes' has no min/max bounds). No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles. Error handling is implicit (relies on shell exit codes) rather than explicit recovery guidance. Naming is verb-prefixed and clear (agentrelay_status, agentrelay_validate), following pattern:tool conventions.
List the git checkpoints the supervisor made, newest first — validated units you can return to.
Would a unit of roughly N minutes fit before the quota window closes? Exit says go or wait — being cut off mid-unit leaves half-done, unvalidated work.
Build (or refresh) a knowledge graph of the project with graphify, so you can query its structure. Default is the free AST-only build (no tokens); set deep:true for the richer LLM extraction (spends tokens). Returns where graph.html / GRAPH_REPORT.md / graph.json were written. Routed through agentrelay, so if graphify is not installed it says exactly how to add it.
Ask a natural-language question about the project against the knowledge graph built by agentrelay_graph — e.g. "how does the guard decide to block a call?". Much cheaper than re-reading files. Needs graphify installed and a graph already built; says so plainly if either is missing.
The morning report: what the run actually did, every line labelled VERIFIED (a command decided it), RECORDED (observed), or ESTIMATED (a guess). Nothing the model merely claimed appears as fact.
No output schemas documented. LLMs cannot predict return structure (fields, types, pagination). Tools return raw shell output as unstructured text.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Tools are marked READ_ONLY or WRITE in metadata but not in MCP schema, forcing LLMs to infer safety from descriptions.
Parameter 'minutes' in agentrelay_fits lacks numeric bounds. Description says 'default 30' but no min/max specified. LLMs could pass negative or absurdly large values.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
The workspace's supervision status in plain words: which guards are armed and whether the blocking hook is installed, the spending cap, isolation, and whether the mission is waiting on a person.
Run the workspace's configured validators (tests/build/lint) as evidence and report PASS/FAIL per check. This is read-only — it does not commit. A unit is 'done' only if these pass.
Parameter 'budget' in agentrelay_graph_query lacks numeric bounds. No min/max specified; LLMs could request 1M tokens, causing runaway costs.
Error handling is implicit (shell exit codes). No explicit error recovery guidance. If agentrelay binary is missing, error message says 'engine not installed' but does not guide LLM on next steps (e.g., 'Install agentrelay via: brew install agentrelay').