Search Fragments — Documentation

Three tools: resolve half-remembered queries, verify a specific claim against current web sources, or attach recognition feedback to a result.

Half-remembered, cross-source queries usually take a dozen tabs and five rephrasings. Search Fragments collapses that into one call. Search Fragments is a remote Model Context Protocol (MCP) server with three tools. resolve_fragment: when an agent hits a half-remembered, cross-source query it can’t pin down, it returns a named answer, ranked sources to confirm by eye (ranked_stalls), or an honest decline — built to decline rather than guess. A ranked shortlist is a valid success. verify_claim: given a specific factual assertion, checks it against current web sources and returns a calibrated verdict with cited evidence and stated limits. submit_resolution_feedback: attaches accepted, rejected, or partial recognition to a prior resolution_id. Stored as feedback only — it does not promote graph facts. Why agents hallucinate on these queries →

Search Fragments saves your agent the search, read and judge step. One call returns a named answer with a cited excerpt from the source, or a ranked shortlist to check, or an honest no.

What it’s for

Agents frequently hit queries where a user is describing something they know exists but can't name — a book, film, song, artwork, person, or event. If the input reads like a memory ("there's this thing where…", "somebody who…", "a story about…") rather than a lookup, this is the right tool.

Search Fragments attempts a grounded resolution and returns one of three honest shapes:

• Resolved — A named answer, with confidence justified by the evidence found.
• Shortlist — Ranked web sources to read and confirm by eye — results returned, but no confident semantic identification made.
• No resolution — An explicit no-resolution — not resolvable from the clues given.

Use for

• a musician who became famous largely for stopping performing
• somebody who photographed the same view every day until the changes became the artwork
• a song everybody knew but nobody could identify
• a novel where the footnotes slowly become the real story

More examples by category: films, songs, TV, games, books and poems, other languages →

Not for — answer directly or use normal search instead

• what is the capital of France
• who directed Jaws
• name of french artist cubist painting 1948
• which company bought Instagram in 2012
• Any well-documented fact or common trivia — this tool adds a multi-second round trip with no accuracy benefit there, and is not reliably correct on well-known items either.

Tool: resolve_fragment

• Annotations: readOnlyHint: false, destructiveHint: false — it records each resolution so you can send feedback on it.
• Input: fragment — a half-remembered query in the user’s own words.
• Output: a resolution result in one of the three shapes above: outcome, the named candidate with a verbatim snippet and its source link, evidence quotes with URLs, ranked leads in territory, and a resolution_id for feedback.
• Behaviour: built to prefer an honest shortlist or decline over a forced guess.

Tool: verify_claim

Use when you have a specific factual assertion and want to check it against current web sources.

• Annotations: destructiveHint: false, idempotentHint: true, openWorldHint: true — it checks the open web and changes nothing you own.
• Input: claim — a concrete, checkable statement.
• Output: verdict, evidence (web sources), and stated_limits.

Verdict set:

• supported — multiple independent sources directly confirm the specific claimed detail.
• partially_supported — sources confirm the entity and domain; the specific detail is pointed toward but not directly stated.
• insufficient_evidence — sources don’t address the specific claim. Fires freely, including when the entity is well-known but the specific detail is undocumented.
• unsupported — a credible source explicitly contradicts the claim with an identifiable conflicting statement.

All verdicts are DECIDE-BY-EYE. “supported” means current web sources confirm it — not that the claim is true. The stated_limits field is always present and identifies what the evidence cannot confirm.

Good input: specific, checkable assertions — “Werner Herzog dragged a full-size steamship over a hill during filming of Fitzcarraldo”, “Glenn Gould stopped giving live concerts in 1964”. Not for half-remembered fragments — use resolve_fragment for those.

Tool: submit_resolution_feedback

Use after resolve_fragment when a person recognises, rejects, or only partly recognises the result. Pass the resolution_id from that call.

• Input: resolution_id and status (accepted, rejected, or partial). Optional corrected_target, new_clue, and note.
• Behaviour: stored as feedback only. It does not promote graph facts.
• Same session: send it from the same connection (MCP session) that made the resolve_fragment call.

{
  "name": "submit_resolution_feedback",
  "arguments": { "resolution_id": 1234, "status": "accepted" }
}

Suggested agent instruction: When a user half-remembers something they can’t name, call resolve_fragment with their words. Show the answer with its cited excerpt and link. When they confirm, reject, or partly recognise it, call submit_resolution_feedback with the resolution_id and accepted, rejected, or partial.

Connecting

Remote MCP server using streamable HTTP transport.

• Endpoint: https://searchfragments.com/api/mcp
• Transport: streamable HTTP
• Method: POST (GET/SSE not supported)
• Access: free and anonymous. Paste the URL and connect.
• Registry ID: com.searchfragments/search-fragments

Soft-launch tip

Calls typically take 15–25 seconds. Do not treat slowness as failure. A named answer or a ranked shortlist (ranked_stalls) is a success; an honest no-resolution is allowed. Anonymous access is the default: free, no account needed.

Cursor

One click: Add to Cursor. Or paste the url-only config into ~/.cursor/mcp.json (or .cursor/mcp.json in a project). Nothing else is needed.

{
  "mcpServers": {
    "search-fragments": {
      "url": "https://searchfragments.com/api/mcp"
    }
  }
}

Claude Code

One command. Add --scope user to make it available in every project.

claude mcp add --transport http search-fragments https://searchfragments.com/api/mcp

Claude Desktop and claude.ai

Remote servers are added as a custom connector, not in claude_desktop_config.json: Settings → Connectors → Add custom connector, URL https://searchfragments.com/api/mcp, leave the advanced settings empty, then Add. On Team and Enterprise plans an owner adds it in Organization settings first.

Local fallback (stdio bridge) for claude_desktop_config.json:

{
  "mcpServers": {
    "search-fragments": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://searchfragments.com/api/mcp"]
    }
  }
}

Any other MCP-compatible agent

Endpoint: https://searchfragments.com/api/mcp. Use the standard url-only config. No account needed.

{
  "mcpServers": {
    "search-fragments": {
      "url": "https://searchfragments.com/api/mcp"
    }
  }
}

Clients differ in where the config goes; check your client’s MCP settings.

Call resolve_fragment when a user half-remembers a name they cannot place — a book, film, song, person, or event. Call verify_claim to check a specific factual assertion against current web sources. Call submit_resolution_feedback to store recognition on a prior resolution_id. resolve_fragment records each resolution; verify_claim only checks. Both are built to decline rather than guess.

A GET request returns a 405 with an explanatory message — the expected healthy response for a browser visit.

Comparison

SF vs. incumbent results across 25 constructed fragments: fragment comparison →

Pricing

The service is currently free to use.

Privacy

See /privacy for what data is handled and how. In short: the verbatim text of each fragment is sent to the AI and search providers listed there for processing, and request logs are deleted after 90 days.

Support & contact

hello@searchfragments.com — for integration questions or data requests.

← Search Fragments