Embeds a query, vector-searches raw content, and hands an agent whatever excerpts matched — with no guarantee any of it was reviewed, current or in scope for that caller.
Agents get facts.
Not raw documents.
Every MCP tool call is authenticated, scoped to what the caller is authorised to see, and returns a cited, confidence-scored answer—never an unscoped document dump. Thirteen tools are live today.
01 / THE GAPRetrieval is not
Retrieval is not
the same as grounding.
Structured lookup only. An autonomous caller can retrieve a fact or a skill, but only if it has been grounding-confirmed by a human curator — never a proposed or unconfirmed one.
Confidence, freshness and a provenance summary travel with every result, so an agent — or the person reviewing what it did — can verify the claim without a separate lookup.
02 / THE TOOL SURFACEThirteen tools.
Thirteen tools.
Every one scoped.
Mounted at a single MCP endpoint. Every tool listed here is live in the codebase today, not a roadmap item — grouped by what it does, not alphabetically.
| Tool | Group | What it does | ||
|---|---|---|---|---|
| whoami | Identity | Reports the caller's resolved identity and auth mode, so a failed write explains why. | ||
| search_skills | Skills | Searches published skills by domain or query, scoped to the caller's domain grants. | ||
| get_skill | Skills | Retrieves the full canonical skill document for a published skill. | ||
| invoke_skill | Skills | Executes a published skill through the safe-execution contract; risky steps pause for approval. | ||
| dry_run_skill | Skills | Compiles and risk-classifies a skill without running it, so approval gates are visible first. | ||
| get_run_status | Skills | Polls a skill run until its human-approval gate clears. | ||
| report_outcome | Skills | Reports whether a completed run worked, feeding confidence scoring. | ||
| lookup_facts | Facts & evidence | Structured, cited fact lookup. Only grounding-confirmed facts are returned. | ||
| search_evidence | Facts & evidence | Semantic search over cited evidence, rechecked against the caller's scope on every call. | ||
| get_evidence | Facts & evidence | Fetches a single evidence unit by ID, with a deep link back to its original source. | ||
| get_skill_evidence | Facts & evidence | Shows exactly which source documents ground a specific skill, before you invoke it. | ||
| search_concepts | Vocabulary | Searches the confirmed business-vocabulary layer — named terms like "refund-window." | ||
| get_concept | Vocabulary | Retrieves one confirmed concept with the evidence backing it. | ||
Internal engineering record, reviewed August 2026. Every call is authenticated by API key and re-checked against the caller's domain and permission-tag scope — a caller outside a skill's or fact's scope sees "not found," not a 403 that would confirm the item exists.
03 / SAFE EXECUTIONEvery invocation
Every invocation
can be inspected first.
dry_run_skill compiles a skill to a risk-classified action plan and shows every step's classification—without creating a run. See the approval gate before you trigger it.
invoke_skill routes through the same safety classifier as every other execution path. A step classified as requiring approval pauses the run at "awaiting approval" until a person clears it.
report_outcome feeds the real business result—success, partial or failed—back into the confidence score the next caller sees.
Point your agent
at governed facts.
Tell us which agent needs structured, cited answers instead of raw document access. We will scope the domains and permission tags around it.
Start a scope →