Consensus

Required before citing a search result. Retrieves the full paper record for an `id` returned by `search`, including abstract, authors, journal metadata, citation count, DOI when available, and the canonical Consensus paper URL.

fetch

Keyword search over 400M+ peer-reviewed academic papers. Returns JSON results with `id`, `title`, `text`, and `url`. You MUST call `fetch(id)` on any result before citing it. To request filters, include tokens in the query text: `year:2020-2026`, `date:2020-03` or `date:2020-03-2022-06`, `study:rct`, `study:meta-analysis`, `study:systematic-review`, `human:true`, `domain:med,bio`, `sjr:1` (Q1 journals only) or `sjr:1-2` (Q1 and Q2 journals; SCImago quartile, 1 is best; only when the user asks for journal quality). Valid domain codes: med, bio, cs, chem, psych, phys, mat, eng, env, bus, econ, math, poli, agri, edu, soc, geol, geog, hist, art, philo, law, ling.

search

Add a new interaction to an existing research thread. Only for a genuinely new follow-up the user raised after the previous interaction's results were delivered. Never chain this onto your own create_thread call to work through parts of a single request — put the whole request in the create_thread brief instead. Call this only once get_thread reports the thread 'idle' or 'failed'. While an earlier interaction is still running this returns an error: keep polling get_thread until status is no longer 'running', then call add_to_thread again. The new interaction runs asynchronously. After calling, wait at least 30 seconds (use a sleep/timeout) before each get_thread poll. Polling faster returns cached 'running' results and does not speed up the agent.

Create a new research thread and dispatch the Consensus research agent. Runs Consensus's specialized scientific research agent: it plans and runs many searches, reads full text, follows citations forward and backward, analyzes and reflects on the evidence, and returns a cited synthesis. Use it for any research question that needs synthesis or analysis rather than a single lookup: what the evidence says, comparisons, mechanisms, literature reviews, or multi-step research tasks. For a single lookup of papers matching a scientific query, use `search`. Also use this to list the user's library collections, create or rename one, or save papers into one. This dispatches an autonomous research agent that reasons and executes many steps on its own. Delegate the entire request in a single call: pass a full brief covering every part of what the user asked, including ordered sub-tasks ('identify A, then find B related to A'). Do not decompose the request into successive create_thread/add_to_thread calls, and do not do intermediate research yourself to hand it a narrower question — that yields worse results and costs several minutes per hop. After calling, wait at least 30 seconds (use a sleep/timeout) before each get_thread poll. Pro results typically finish within ~3 minutes; Deep results may take up to ~10 minutes. Polling faster returns cached 'running' results and does not speed up the agent.

List the signed-in user's research threads. Requires a signed-in Consensus account. Returns thread metadata including a deep-link URL.

Poll a research thread and return the grounded answer when ready. The returned status is 'running' while the agent works and 'idle' when a final answer is available. Wait at least 30 seconds (use a sleep/timeout) between get_thread calls. Pro results typically finish within ~3 minutes; Deep results may take up to ~10 minutes. Calling more frequently than every 30 seconds returns cached 'running' results and does not speed up the agent. When status is 'idle', write the interaction's `document` field to a file or the client's artifact/canvas/side-panel surface — do not paste it inline in chat and do not paraphrase it. Every inline citation marker, [N.M] (quote M of paper N) or a bare [N], must be preserved exactly; they are already hyperlinked to the cited papers.