Consensus
by Consensus
Search 220M+ peer-reviewed papers from your AI assistant. One tool, read-only, free without an account. Search returns real abstracts and metadata — not the consensus meter verdict — and retracted papers are flagged and excluded from analyses.
Verified connector
Listed by Anthropic as a partner connector in its Connectors Directory.
Connection checked by Agentman on .
Anthropic states this reflects the level of review a connector received, not a security audit.
Connect Consensus via MCP
https://mcp.consensus.app/mcpWorks in any MCP-compatible client. In Agentman Studio it is one click — no config file to edit.
Consensus Tools & Capabilities (1)
Tool names from Anthropic's directory listing. This server requires sign-in, so we could not read tool descriptions or parameter schemas.
Limits
- Consensus says a model can misread a paper. It states that of three hallucination categories only misread sources is possible in its architecture, and that no AI is perfect (Consensus help centre, fetched 2026-08-19). Real citations do not guarantee a correct reading of them.
- The consensus meter has a published error rate of 10%. Consensus calculates its classifier gets results wrong one time in ten, more often with double negatives and convoluted language, and describes the feature as beta (Consensus meter blog post, fetched 2026-08-19).
- The meter ignores study quality. Consensus states each claim counts equally regardless of whether it comes from a meta-analysis or a single case report. A journal-quartile filter is not a quality judgement.
- The meter can miss the question's nuance. Consensus's own example is a query about adults being answered by a finding about children.
- Coverage is not comprehensive and Consensus says so. It describes 220 million+ papers as significant coverage but not all research, and results as a snapshot rather than a full view of the literature.
- Full text is not always available. Consensus states availability varies by paper and that it falls back to metadata and the abstract otherwise; where full text is used, the article may still sit behind a paywall.
- The preprint default is undocumented.
exclude_preprintsexists, but the docs do not say which way it falls when unset. Set it explicitly. - Retracted papers have no documented flag in the MCP response. Consensus excludes them from analyses and badges them in-product, but no retraction field appears in the documented search response fields, and the docs do not say whether a retracted paper can still surface as a raw search result.
- Search volume is capped by plan — 30 a month free, 250 on Pro, 1,000 on Deep and Teams — and the docs do not say whether MCP searches share that count with web app usage.
- Result depth is plan-gated too. No account returns 3 papers;
study_typeandtakeawayrequire Pro or above; DOI is documented as enterprise. - The terms disclaim accuracy and medical use. Consensus provides the Services as-is without warranty of accuracy or completeness, and states it does not provide medical advice and that its information is not intended for diagnosis, evaluation or treatment.
- Reuse rights are narrow. The terms permit personal-use copying only and prohibit other reproduction, publication, distribution, modification, derivative works and sale of site content.
- We could not verify tool behaviour ourselves. The endpoint returned HTTP 401 to an anonymous handshake on 2026-08-19, so no live schemas and no safety annotations were readable. The tool names and response fields come from Anthropic's directory and Consensus's documentation.
- Prompts are unconfirmed. Anthropic's snapshot lists no prompt names, and the gated endpoint blocked our own listing call, so we cannot say whether the server serves any.
- The
profilescope is unexplained. It appears on the resource descriptor but not in the authorization server's metadata or the challenge header, and Consensus's docs do not mention it.
Frequently asked questions
What does the Consensus search tool actually return?
Retrieved paper records, not a verdict. Consensus documents each result as title, authors, abstract, journal, year, citation count and a link, plus a total and the original query. The consensus meter that summarises whether the literature agrees is a web-app feature and is not among the documented MCP response fields. One generated field exists — a per-paper key takeaway, available on Pro plans and above.
How does Consensus handle retracted papers?
It flags them and keeps them out of its analyses. Consensus's help centre states retracted papers carry a visible retracted badge, and that they are never used in its analyses or summaries. That is a stronger stance than an opt-out filter, though the flag itself is described as a product marking and does not appear among the documented MCP search response fields.
Is the Consensus MCP server free, and does it use up my searches?
It is usable with no account at all, and yes, it draws on your Consensus plan quota. Consensus documents an unlimited-search tier without an account returning 3 papers per search, a free account at 10 papers and 30 searches a month, and Pro at 20 papers and 250 searches. Your Claude or ChatGPT subscription has no effect.
Does the Consensus MCP server only have one tool?
Anthropic's directory lists one; Consensus documents two. The second is a fetch tool that retrieves full details for one paper, and Consensus states it is currently available for ChatGPT only, with other MCP clients seeing just the search tool. So the directory's single-tool listing is accurate for Claude and most other clients, and incomplete as a description of the server.
Can Consensus misread a paper, and does it say so?
Yes to both. Consensus states that of three hallucination categories only misread sources is possible in its design, because it searches before summarising, and it works to minimise that. Separately it puts a figure on the consensus meter: it calculates the classifier gets results wrong 10% of the time, more often with double-negatives and convoluted phrasing.
What corpus does Consensus search?
Over 220 million peer-reviewed papers, updated weekly. Consensus names four sources: Semantic Scholar, OpenAlex, its own crawl of the scholarly web, and publisher partnerships providing full text. PubMed is covered in its entirety. A separate medical mode narrows to roughly 8 million papers and 50,000 clinical guidelines from the top 1,000 medical journals.
Can I reuse the abstracts and summaries Consensus returns?
Consensus's terms grant only limited personal-use rights. They permit downloading or copying site content for personal use with no transfer of rights, and state that other reproduction, publication, distribution, modification, derivative works or sale is prohibited. The terms also disclaim warranties as to accuracy and state Consensus does not provide medical advice. Read them yourself before republishing; this is not legal advice.
Sources
- Consensus MCP documentation — https://docs.consensus.app/docs/mcp (fetched 2026-08-19 as Markdown via the documented
.mdsuffix; page metadata dates it 2026-06-22 — the source for the plan table, tool parameters, response fields and the ChatGPT-onlyfetchnote) · retrieved 2026-08-19 - Consensus API reference, search endpoint — https://docs.consensus.app/reference/v1_search.md (fetched 2026-08-19; OpenAPI 3.1.0, API version 1.0 — the source for the full per-paper field schema) · retrieved 2026-08-19
- Consensus documentation index — https://docs.consensus.app/llms.txt (fetched 2026-08-19; lists three pages, the MCP guide and two search endpoints) · retrieved 2026-08-19
- Consensus help centre, research database and retraction policy — https://help.consensus.app/en/articles/10055108-consensus-research-database (fetched 2026-08-19) · retrieved 2026-08-19
- Consensus help centre, responsible AI and limitations — https://help.consensus.app/en/articles/10083353-responsible-ai-limitations (fetched 2026-08-19; dated 5 June 2026) · retrieved 2026-08-19
- Consensus Meter: Guardrails and Limitations — https://consensus.app/home/blog/consensus-meter/ (fetched 2026-08-19; dated 31 January 2023 — the source for the 10% misclassification figure) · retrieved 2026-08-19
- Consensus terms of service — https://consensus.app/home/terms-of-service/ (fetched 2026-08-19; dated 25 June 2025) · retrieved 2026-08-19
docs.consensus.app/robots.txtisAllow: /with targeted disallows for edit, login and internal API paths, and points atllms.txt; it carries noContent-Signaldirectives.consensus.app/robots.txtdisallows/blog/and sets a crawl delay of 1, while explicitly allowing OAI-SearchBot, ChatGPT-User, PerplexityBot and Perplexity-User (both fetched 2026-08-19) · retrieved 2026-08-19- Live RFC 9728 protected-resource descriptor — https://mcp.consensus.app/.well-known/oauth-protected-resource and its path-insert variant (both fetched 2026-08-19, byte-identical at 168 bytes; the path-append form returns 401) · retrieved 2026-08-19
- Live RFC 8414 authorization-server metadata — https://mcp.consensus.app/.well-known/oauth-authorization-server and its path-insert variant (both fetched 2026-08-19, byte-identical;
scopes_supportedis["search"]only) · retrieved 2026-08-19 - Live anonymous probe — 2026-08-19: an
initializerequest to the endpoint returns HTTP 401 withwww-authenticate: Bearer resource_metadata="…", scope="search". No tool was called and no authentication was attempted · retrieved 2026-08-19 - Anthropic connector directory — https://claude.ai/directory/65247229-f0c7-49df-9044-fcbb8b3894c6 (snapshot 2026-08-16; permissions label "Read Only", one tool name, no prompt names) · retrieved 2026-08-16
Use in Agentman
Connect once and your agents call these tools on their own — on a schedule, in a workflow, with nobody at the keyboard.
Open in Agentman StudioServer Info
- Category
- Data & Research
- Developer
- Consensus
- Tools
- 1
- Domain
- mcp.consensus.app
Using Claude Desktop or another MCP client? Setup docs — the connection URL above works anywhere.