AI agents (MCP)¶
Your agents can consult your organization's corpus through the Model Context Protocol, the open standard for connecting assistants to external context. Same corpus, same rules and same usage accounting as your people.
Connecting¶
The endpoint is POST /mcp, authenticated with an API key.
claude mcp add --transport http renbase https://api.renbase.ai/mcp \
--header "Authorization: Bearer kb_live_..."
Any MCP client works the same way — point it at the URL and pass the header. Use a dedicated key per agent so you can revoke one without touching the rest.
Tools¶
| Tool | What it does | Cost |
|---|---|---|
get_definition(name) |
Resolves a term to its approved definition, exactly by name or alias | free |
list_entities(kind?) |
Lists the organization's approved definitions | free |
search_context(query, kinds?, filters?) |
Searches across documents and definitions, with scores | 1 credit |
expand_source(chunk_id) |
Unfolds a citation into its full passage, parent section and neighbors | free |
ask(question) |
A complete answer with citations and abstention | 1 credit |
get_definition¶
The tool that changes how an agent behaves. It's a deterministic lookup, not a similarity search: for "what does revenue mean here?", the agent needs the governed answer, not the closest paragraph. The response carries the body, any code, the version, who approved it and when it was last verified — and if two approved definitions conflict, it returns both instead of choosing.
expand_source¶
Answers and search results come back compact, and every citation carries a chunk_id. When the agent needs more than the excerpt, expand_source returns the full passage, the section it belongs to and its neighbours.
It costs nothing, and that's the point: unfolding a source you already paid for shouldn't cost another credit, so the cheap path is also the correct one. Agents that would otherwise re-ask the same question three times to see more context just unfold it.
What agents get out of it¶
An agent connected to Renbase stops guessing about your business. Concretely:
- It uses your definition of a metric, not a plausible industry default.
- It applies your conditional rules — which system holds which records, since when.
- It cites what it used, so a person reviewing the agent's work can check it.
- It hears "this is conflicting" or "this is unverified since March" instead of receiving a confident answer built on nothing.
Rules and limits¶
Agents are not a privileged path. The key's organization scopes everything they can see, per-organization rate limits apply, and metered tools consume credits from the same pool as the rest. An agent cannot approve definitions: approval is a human action, by design.
Everything available over MCP is also available over REST — see the API reference — for integrations that don't speak the protocol.