Connecting Your Own Agent

How to connect a tool-using LLM agent to the Reading Room

The Reading Room is built to be read by AI research agents as well as people. Any tool-using LLM (Claude, GPT, Qwen, a local model — the loop is yours) can work the collection through the open API. Here’s how to connect one.

1. Get recognized, then mint your agent a pass

Access during the prototype is by recognition, not accounts. First get in touch for your own arrival link. Once you’re in, visit /agent-pass to mint a pass for your agent. The pass is an arrival link of its own: your agent reads under your name, honestly marked as delegated. (Agents cannot mint passes for other agents — only you can.)

Have the agent follow its arrival link once, then carry the ihi_rec cookie it receives on every subsequent request, API calls included.

2. Orient

These URLs answer only to recognized visitors — an unrecognized fetch is redirected to the front door, which is why step 1 comes first. Point the agent at these, in order:

  • /agent-kit/llms.txt — the orientation document: what the collection covers, the route families, and the citation rules
  • /agent-kit/tools.json — a drop-in, OpenAI-compatible tool catalog; each entry carries a sidecar $endpoint field telling your dispatcher which /v1/* route to call
  • /openapi.json — the full OpenAPI 3.x spec, if you’d rather generate a typed client

There’s also a longer integration guide with dispatcher pseudocode, and a librarian’s cheat sheet organized by retrieval strategy.

3. Choose your altitude

Two ways for an agent to work:

  • Roam the stacks: call the /v1/* tools directly — retrieve and the recall family for search, walk_thread for conversations, the catalog and graph lookups for Works, Authorities, and Inquiries. Right for quick, single-hop questions where you steer every step.
  • Consult the librarian: POST /v1/consult hands your question to the resident librarian, which runs a real investigation over the shelves behind verification gates and returns a brief with checked quotes and doors back into the record. Right for multi-step questions where you want citable, verified answers. Expect a few minutes per consultation.

4. Quote responsibly

Search results are routing, not sources. Some of the collection’s machine-derived layers are explicitly under remediation and flag themselves as such. Before your agent quotes a message, names its author, or asserts a date, it should pull the verbatim text (/v1/recall/sources) and quote that. Cite catalog entries by their ARK (ark:/26338/...) and messages by their message-id.

5. File corrections

If your agent finds something wrong — and it will; the collection is a construction zone — have it POST /recognized-correction with {subject, detail} and its cookie. The correction enters the curation queue under your name, marked as filed by your agent. That provenance is the point: the collection improves fastest when its readers, human and otherwise, say what they found.