Search and research MCP for knowledge work.
One server to search the web and specialised indexes built for knowledge work, and run multi-step research with cited outputs, from Claude, Codex or any agent. Read the docs.
Hermespiand more…Connect
Claude
Settings → Connectors → Add custom connector. Paste the endpoint and sign in with Valyu.
https://mcp.valyu.ai/mcp
ChatGPT
Settings → Connectors → Create. Paste the endpoint and sign in with Valyu.
https://mcp.valyu.ai/mcp
Claude Code
claude mcp add --transport http valyu https://mcp.valyu.ai/mcp
Codex, Cursor and others
Any client that speaks Streamable HTTP. Sign in with Valyu when prompted, or put an API key on the URL (an Authorization: Bearer header works too):
{ "mcpServers": { "valyu": { "type": "http", "url": "https://mcp.valyu.ai/mcp?valyuApiKey=<VALYU_API_KEY>" } } }
Hermes, pi and terminal agents
Point the agent at the endpoint; it signs in on first use. Keys live at platform.valyu.ai.
Tools
Any question that needs current information, primary sources or citations. Returns ranked results with content, URLs and relevance; scope with included_sources for precision, leave it unscoped for breadth.
Natural language query.
ADVANCED, and a hard filter: everything not named is excluded, so a wrong value returns less than no filter at all.
ADVANCED, but the safe one: it only reorders, so it cannot empty the result set.
ADVANCED.
How many distinct documents to return.
Maximum characters returned per result (500-100000).
ADVANCED, and the option most likely to return nothing by accident.
ADVANCED.
Return figures extracted from each result, as links.
ADVANCED.
Call before a scoped search to see which specialist datasets exist for a topic. Returns dataset ids, what each covers, and whether the plan includes it.
Describe the data you need in plain language, e.g. "peer-reviewed medical literature", "FDA drug labels", "insider transactions", "CPI inflation".
Annotate each dataset with whether the current plan includes it.
v1 option: only list datasets in this category id.
Read one or more pages in full. Returns clean markdown, optionally summarised to an instruction.
1-10 absolute http(s) URLs to fetch.
true for a general summary, or a string instruction describing what to extract, e.g. "only the methodology and results".
"normal" is fastest; "high" renders JavaScript and is slower but succeeds on SPAs and difficult pages; "auto" picks per URL.
Also return a full-page screenshot of each URL, as a link.
Maximum characters returned per page (500-200000).
Only when the user wants a written deliverable: a report, literature review, due diligence. Runs for minutes to hours as a research agent; returns a task id immediately.
The deepresearch brief.
Depth/cost tier.
Block for up to this many seconds waiting for completion before returning a task id.
Natural-language guidance for the report structure, e.g. "a one-page memo with a risks table" or "bullet points, no preamble".
Natural-language guidance for how to run the research, e.g. "prioritise primary sources and SEC filings over news".
ADVANCED, and a hard filter.
ADVANCED.
ADVANCED, but the safe one: it only reorders and cannot starve the report.
ADVANCED.
ADVANCED.
ADVANCED, and the easiest way to starve a report by accident.
ADVANCED.
Up to 10 URLs the agent must read and use as context (e.g. a filing, a trial page, a competitor's site).
Up to 3 earlier task ids whose reports become context for this one - for follow-ups and updates.
Up to 10 files (as base64 data URLs) for the agent to analyse.
Also produce a PDF of the report (a download link is returned with the report).
JSON Schema for structured JSON output instead of a markdown report.
Up to 10 files to generate from the report, e.g. "an Excel sheet of every trial with phase and sponsor" or {"type": "pptx", "description": "10-slide summary"}.
Enable or disable optional agent tools.
Human-in-the-loop checkpoints.
Up to 5 external MCP servers whose tools the deepresearch agent may call (e.g. an internal data source).
HTTPS URL to POST to when the task completes (HMAC-SHA256 signed).
Email the finished report link to this address (must belong to the user's organisation).
Arbitrary key/value tags stored with the task.
Beta: run a saved Valyu workflow (prompt template) instead of a free-form query.
Values for the workflow's template variables, keyed by variable key.
Pin a specific workflow version; defaults to the current one.
Check a running task, collect the finished report with citations, or, with no task id, list recent tasks.
The task id returned by valyu_deepresearch.
Listing only: how many recent tasks to return (default 20).
0 (default) returns the current status immediately - right for a conversation: report it and let the user come back later.
Include the citation list with deep links into each source (completed tasks only).
Redirect a running task, answer a checkpoint it is waiting on, toggle the share link, or cancel or delete it.
The task id returned by valyu_deepresearch (or listed by valyu_deepresearch_status).
"steer", "cancel", "respond", "share" or "delete" - see the tool description.
The steering instruction, in plain language.
The checkpoint id shown by valyu_deepresearch_status.
The checkpoint answer, in the shape valyu_deepresearch_status specified for that checkpoint type.
Required when action is "share": true publishes a share link, false removes it.
Sources
The live web plus academic, financial, clinical, patent, economic, legal and security datasets, growing every month. Browse every data source.