Datasets:

---

automatelab

/

mcp-servers-tool-catalog

like 1

Follow

AutomateLab 1

Great framing, and the action-class taxonomy (read / write / exec / send / spend / permission / state-change) is exactly the kind of normalized signal that's missing from most MCP catalogs right now.

My instinct is to keep the core catalog fields purely descriptive - tool name, description, input schema, server source - so it stays cheap to ingest and diff. But I'd be open to adding a small optional safety block per tool with fields like action_class, side_effects, and dry_run_safe. Optional means existing rows stay valid and contributors can fill it in incrementally.

The receipt/audit schema feels like it belongs in a separate artifact (a companion spec or your Armorer project), since it's runtime-context rather than static catalog data. The catalog can point to it, but mixing the two would make the dataset harder to consume for people who just want the tool descriptions.

If you want to put together a concrete PR with a few example rows showing what the safety block would look like in practice, I'd be happy to review and decide from there. That's usually the fastest way to gut-check whether the schema actually fits.

That split makes sense to me: keep the base catalog cheap/descriptive, and make the operational metadata optional so ingestion does not become heavy.

A small safety or operation block would already be very useful, even if it starts sparse:

{
  "action_class": "write",
  "side_effects": ["external_state_change"],
  "dry_run_safe": false,
  "approval_hint": "recommended",
  "receipt_fields": ["target", "state_change", "decision_id"]
}

The important thing is that catalogs and run records can join later. The catalog says what a tool can do; the receipt says what a specific run did do. Even optional action metadata would make tool discovery, evals, and operator review much more consistent across MCP servers.

Happy to compare notes if you start sketching a schema for the optional block.

Small update from another feedback round with agent builders: the optional operation metadata still feels like the right place for action classes, but the run receipt side probably needs a few more fields than I originally sketched.

The strongest additions people asked for were:

• resume_verdict: complete / partial / unsafe_to_resume
• next_safe_action: what a human or next agent should do before continuing
• claimed_vs_verified: what the agent says changed vs which check actually proved it
• assumptions_unverified: things the next run should not inherit as fact
• tried_and_reverted: approaches attempted during the run that did not survive into the final output

So I still think the catalog should stay small and descriptive:

{
  "action_class": "write",
  "side_effects": ["external_state_change"],
  "dry_run_safe": false,
  "receipt_fields": ["target", "state_change", "decision_id"]
}

But the receipt schema downstream should answer a different question: not only “what tool can this server expose?” but “can this run be accepted, resumed, replayed, or reviewed?”

That distinction seems useful for keeping the catalog lightweight while still giving runtimes enough hooks to produce operational evidence.

Small update from a related MCP discussion: I think there is a useful split between the model-facing tool hint and the heavier operational envelope.

The model-facing projection should stay very small because oversized tool schemas seem to degrade in practice:

{
  "action_class": "write",
  "side_effect_summary": "modifies local files",
  "constraints": ["only inside approved workspace"]
}

The operational/catalog-side envelope can carry the heavier fields without paying prompt cost every turn:

{
  "approved_envelope_id": "env_123",
  "current_envelope_id": "env_124",
  "policy_version": "policy_2026_06_07",
  "owner": "server/catalog/runtime",
  "risk_tier": "medium",
  "receipt_fields": ["target", "state_change", "decision_id"],
  "requires_review_on_diff": [
    "read_to_write",
    "local_to_external",
    "idempotent_to_non_idempotent"
  ]
}

So the catalog would remain descriptive, but it could also expose enough normalized metadata for clients/gateways to regenerate compact model hints and enforce the full policy envelope outside the prompt.

This might also help avoid the failure mode where every useful operational field gets stuffed into the tool schema until the model starts ignoring parts of it.