AI agents (MCP)

Coding agents are fast and have no memory of your organization. They will happily rebuild something the team already tried, or reopen a question that was settled two years ago in a document nobody thinks to mention. Connecting Rationale over MCP gives an agent access to the reasoning behind your systems - and, importantly, to the qualifiers that come with it.

What an agent can do with it

Rationale runs a Model Context Protocol server, so any MCP-capable host can use these tools directly:

Check a proposal - the one worth connecting for. Describe a change you are about to make and get back the decisions that already cover that ground, alternatives your team considered and rejected with the reason they recorded, assumptions the change depends on and whether they still hold, and any direct contradictions.
Find decisions - search the map for what has been decided on a topic.
Ask- a synthesized, cited answer to a "why do we…" question.
Explain a decision - the full record: the problem, the choice, the rationale, the alternatives, the evidence, and its history.
Find related decisions - walk out to what a decision supersedes, depends on, or conflicts with.
Who decided - the people a source document named, quoted as recorded. Attribution, not an expertise ranking.
Workspace context - how much your map actually contains. An agent cannot see a sparse dashboard and calibrate, so it needs to be told.

The server also supplies three ready-made workflows your host can surface as commands: pre-change review, incident context, and onboard to an area. Each one scripts the right sequence of the tools above.

Connecting Claude Code

Open Settings → API access in Rationale and choose Connect agent. Pick a workspace, confirm what you are connecting, and you will be given a command to paste:

claude mcp add --transport http rationale https://api.therationale.ai/mcp --header "Authorization: Bearer rk_..."

That is the whole setup - no OAuth flow, no browser round trip on the machine running the agent, which means it also works on a CI runner or a server where no one can click anything. Once the connection is confirmed, you can move the credential into an environment variable; Claude Code expands ${VAR} inside header values.

Claude Code is the host we have tested against. Any MCP host that accepts a static bearer header should work the same way, since that is all the server requires.

What the agent can reach

One workspace. The workspace comes from the credential itself - it is not something a tool call can name - so an agent cannot reach another workspace even if it tries. To connect a second, run the flow again.
Read-only. Agents can ask; they cannot add, edit, or delete anything in your map.
Reasoning, not files. The tools return decisions, assumptions, and citations. Rationale is not a document search tool, and it does not hand agents raw excerpts of your files to page through.

Why the answers still hedge

This is the part we care most about, and it is the reason connecting an agent to a knowledge base is not automatically a good idea.

Rationale never returns a bare answer. Every result says what it rests on - one source or several, verified by a person or not, current or stale, contradicted by something else on record - and it says it in the first line, before the answer, not in a footnote. That is deliberate: an agent that reads a confident sentence will write a confident sentence, and a hedged decision restated as settled fact in a pull request is worse than no answer at all.

The same rule applies to what Rationale cannot see. If your map is thin, a search that finds nothing says so explicitly rather than implying the question was never decided. And a proposal check that finds no conflict reports that - nothing on record collides - which is not the same as approval. Rationale has not reviewed your change; it has checked your change against what was written down.

Good to know

Agent connections expire after 90 days. You will see a countdown next to the connection in Settings, and reconnecting takes the same two clicks.
You can revoke a connection at any time from Settings → API access. Revocation takes effect on the next request.
Rationale watches for runaway agents. If a connection starts behaving unusually - a retry loop, a sudden burst of traffic - you get a note in Settings explaining what it saw, with a button to confirm the traffic is expected. If nobody responds and the behavior continues, the connection is paused automatically and you can resume it.
Requests are rate-limited by cost: the tools that read records are generous, and the ones that call a model are tighter. A well-behaved agent will not notice.
Agent usage is billed and attributed like any other AI traffic in your workspace, and shows up per connection in Settings.

What we deliberately did not build

Rationale is a memory server, not an agent framework. It does not plan, delegate to other tools, or coordinate a workflow - your host already does that far better, and a nondeterministic loop sitting between you and an answer makes provenance impossible to reconstruct. Connect Rationale alongside your other MCP servers and let the host orchestrate: they supply the current state of your systems, we supply the reasoning behind them.

API access covers the REST endpoints behind the same credentials. If you want to try this with your team and something is unclear, contact us or write to contact@therationale.ai.