Connect an AI agent

Follow Connect to Claude to configure a custom connector. The hosted endpoint is:

https://api.webcite.co/mcp

Use an Authorization: Bearer YOUR_API_KEY header as described on the connect page. The public profile exposes the API workflows plus webcite_guide; mcp-health reports the live version and toolCount (34 public tools verified 9 October 2026). Call the guide first to select a workflow. Tool availability does not override backend feature gates.

Tools added since the earlier 22-tool profile:

Set the environment variable before starting the server:

export WEBCITE_API_KEY="YOUR_API_KEY"
npx -y webcite-mcp-server

For Claude Code, with the key available in its environment:

claude mcp add webcite -- npx -y webcite-mcp-server

See the package documentation for profiles and configuration. Context and evaluation tools are opt-in and have separate backend requirements.

Hosted upload_file takes a filename and base64 content, up to 20 MB. It cannot read a path on your computer. File paths are supported only by a local MCP server with access to that file.

For a file at a public HTTPS URL, use upload_url. It stores the original bytes (up to 100 MB; 1 credit on successful storage) and returns an owned asset_id, asset_url and source_version_id. It cannot read a local file or a private chat attachment. After an uncertain failure, keep the receipt instead of repeating the upload automatically.

To check every material claim in an uploaded file, call review_document with its existing asset_id. Separate verify_claim calls can omit claims and lose saved progress. The tool returns a job_id immediately. Poll get_document_review_job until the job reaches complete, partial_coverage, credits_exhausted or failed, then page through claims and uncovered source passages with get_document_review.

A completed job means processing finished; it does not prove complete source coverage. Report full coverage only when the saved review has coverage_complete: true and no pending claims. To read a balance on request, or after an unresolved credit refusal, call get_credit_balance.

An agent or workflow builder can use OpenAPI to inspect operation names, parameters, and schemas. Configure API-key authentication with header name x-api-key. Handle non-success status codes and preserve source IDs in later steps.

For documentation context, start with llms.txt. Every page has a .md equivalent; llms-full.txt combines the guides and reference. These formats provide direct access to the text, without requiring a JavaScript browser.

The hosted API uses evidence policy 13. The hosted MCP connector runs the current release; mcp-health reports its live version and tool count. A local MCP server runs the npm package version you installed. Inspect response metadata and the installed client version; saved responses retain their original policy and evidence time.

• verify_claim and verify_claim_stream return readable text plus structuredContent containing the complete backend response, including optional evidence and real IDs.
• verify_batch returns { results: [...] }, preserving item records, binding details and feedback tokens.
• get_citation preserves the API { data: ... } envelope and adds normalized citations and optional final_response from stored metadata (or a verdict-bearing legacy object). The snapshot preserves final evidence before credit settlement, so usage fields can differ from the final response. It does not generate a replacement verdict.
• Text labels unknown and heuristic scores; 0 is displayed as zero. Generated snippets are not presented as publisher quotations.
• Streaming requires a full result followed by done. Partial groups, missing completion, backend errors and accounting errors return partial_result, not raw events presented as a finished verification. Successful stream usage is preserved as stream_usage.

Malformed stored JSON or response shapes return invalid_api_output. Check isError before consuming structured evidence. The package changes do not alter backend feature gates or prove that historical sources were read.

The default profile is public, which exposes the public API workflows and webcite_guide. A local core profile exposes a smaller tool set. Context and evaluation tools are opt-in and require their corresponding backend access and feature gates. Check the actual tool catalog and /mcp-health before configuring a client; a package version alone does not prove the serving backend supports every workflow.

Public tool catalog (34 public tools verified 9 October 2026; check mcp-health for the live version and tool count):

Account-scoped context workflows file owned source versions into folders, analyze those sources, then query or generate against the retained analysis. They use /api/v2/folders/{id}/query and /api/v2/folders/{id}/generate, not a universal /api/v2/context prefix. Use account authentication or a scoped app key with permission to the selected folder; consult the enabled server contract for those routes. Public-profile MCP tools do not grant account access.

Folder citations identify retained source units and spans. Memory notes and generated retrieval queries are context, not evidence. A source not searched or not yet analyzed remains a coverage gap, not proof that a fact is absent. Generation is a separate 5-credit operation on a completed billable result; preserve its actual usage receipt and any unresolved billing state. Do not resubmit after an uncertain outcome merely because no draft reached your client.

On servers implementing retained generation navigation, evidence_refs selects retained source-unit hints. At the default retrieval limit of 16, hints and their mandatory core context keep a separate 32,000-character allowance. Let N be the number of validated retained navigation queries, or 1 when none are present; at most 16 queries are accepted. Ordinary retrieval has a 32,000 × N character allowance after that core, for at most 32,000 + 32,000 × N retained-text characters. Without references, the text limit is 32,000 × N characters. Ordinary retrieval admits up to 16 × N anchors, with a physical-unit allowance of 48 × N plus three times the reference count. The mandatory hint stage keeps its original 48 plus three times the reference count allowance. Each passage contributes at most 2,000 UTF-16 characters. With no navigation queries, N is 1 and the original allowances remain unchanged. These limits do not bound the complete prompt or provider tokens, or guarantee full source coverage. retrieval.evidence_refs reports requested and offered counts and omissions with source_unit_id and a reason: retained_text_budget, physical_unit_budget or quote_outside_passage_window. Offered hints are not necessarily grounded; omissions do not prove absence.

Implicit generation scope includes every authorized filed source in the resolved folder scope; an explicit source_version_ids request is limited to 100 sources. Each preparation, freshness recheck and external folder-answer publication has a 100 MiB (104,857,600-byte) stored-inventory allowance shared across its source representations. It is checked before inventory payload loading and subsequent provider dispatch. Initial preparation checks it before credit reservation; a later freshness recheck can follow an earlier reservation. Exceeding it returns HTTP 409 folder_generation_read_budget_exhausted, without clipping the source scope or returning a missing-evidence answer. Passing the guard does not establish exhaustive coverage. Confirm the serving contract before using these account workflows.