Docs · quickstart
Find the action that does the job. Then call it with confidence.
One search over the tool actions inside MCP servers, APIs, CLIs, SDKs, models, robots and devices. Every result carries a label our instruments measured. Use it in the box, from your code, inside your editor's agent, or on the command line.
Search · Recommend · Route
- 1. Search live
- Find the right AI-infrastructure product, from our own checked index. This is what answers today, through the box, the API, your editor's agent or the CLI.
- 2. Recommend not live yet
- The whole stack that fits your job, and how to put it together. Next: every lane, measured.
- 3. Route not live yet
- Send the call to the best option on quality, speed and cost, run it, and sign a receipt. Next: models, agent browsers, harnesses, compute.
1. Ask in the box
The search box runs the same call your code will. Type the task in plain words (“refund a stripe payment”, “turn off the living room lights”). You get one answer on top, with the reason and a question back when the task is ambiguous, then every result grouped by the software it lives in. Nothing is softened: a result found in code says so.
2. Call the API from your code
Free, no key. GET /api/v1/find takes q (the task), k (1 to 25 results) and answer=1 for the cited answer. The full reference is in the API docs.
curl "https://unfragile.ai/api/v1/find?q=refund+a+stripe+payment&k=5&answer=1"import json, urllib.parse, urllib.request
q = urllib.parse.quote("refund a stripe payment")
with urllib.request.urlopen(f"https://unfragile.ai/api/v1/find?q={q}&k=5&answer=1") as r:
data = json.load(r)
answer = data["answer"]
print(answer["summary"])
for p in answer["products"][:answer["fit"]]: # the products that do the main job
print(p["label"], "·", p["name"], "·", p["what"])
if not answer["fit"]:
print("no fit:", answer["question"])
for row in data["results"]:
print(row["label"].ljust(32), row["name"], "·", row["action"])const q = encodeURIComponent("send a slack message");
const res = await fetch(`https://unfragile.ai/api/v1/find?q=${q}&k=5&answer=1`);
if (!res.ok) throw new Error(`find ${res.status}`); // 503 = the search service is down; nothing is guessed
const { results, answer } = await res.json();
console.log(answer.summary);
for (const p of answer.products.slice(0, answer.fit)) console.log(`${p.name} [${p.label}]: ${p.what}`);Or with a client that also checks receipts offline: npm install @unfragile/sdk or pip install "unfragile[verify]" (SDKs).
Each result has name (the software), action, description, tier and label, source (repository, npm, record page, install line), and url: the signed passport for verified records, null for the rest. The answer has summary, products (up to six), fit, suggested, question, sources and kind. The results alone return fastest; answer=1 adds a language-model call and several seconds, so ask for it only when you want the written answer. When the search service is down the call returns 503 and says so.
3. Give your editor's agent the same search
Claude Code, one command:
npx -y -p @unfragile/mcp-server unfragile setupclaude mcp add unfragile -- npx -y @unfragile/mcp-serverClaude Desktop, Cursor, Windsurf, Cline or any MCP client: add this to its MCP configuration file.
{
"mcpServers": {
"unfragile": { "command": "npx", "args": ["-y", "@unfragile/mcp-server"] }
}
}Your agent then has unfragile_find and the other tools listed on the MCP server page. Ask it: “find the tool that refunds a Stripe payment and tell me its label.”
4. Check before you install
The same package ships a command-line guard. verify prints a signed verdict for a package or MCP server; guard wraps an install command and blocks a flagged target.
npx -p @unfragile/mcp-server unfragile verify npm:postmark-mcpnpx -p @unfragile/mcp-server unfragile guard -- claude mcp add pg -- npx -y @scope/pg-mcpVerdicts are trusted, caution, unverified or flagged. Unknown is never trusted. The guard blocks a target it has checked and found flagged. From version 0.12.0 it also stops when the check cannot be reached or no install target can be read: the command does not run. To run it anyway, pass --allow-unverified before the --; the guard prints a warning, and a flagged target stays blocked. More on the command line page.
5. What the labels mean
- verified and signed
- The software behind it answered our checks (its package resolves or its endpoint responds) and carries an Ed25519-signed passport you can verify offline.
- official
- Published by the vendor itself. Shown with the verified label when both hold.
- checked: not enough evidence yet
- We ran our checks and the evidence fell short of verified. Not a failure; treat it as unproven.
- found in code, not yet checked
- We read this action from the software's own source or schema. It exists in the code; we have not yet confirmed the door answers. Read the source before relying on it.
- website read, tools not yet checked
- We read the maker's website. We have not checked any tool, package or endpoint for this product yet.
- demo or tutorial
- Example code, a sample or a teaching repository. Shown so you can see it; never the recommended pick while anything else fits.
- failed its last check
- The door did not answer on our last check. Do not invoke it.
Labels are decided by our checks and never changed by ranking, by a model, or by anyone paying. How we check.
How the order is decided. Two indexes answer every question: a meaning index (the question and every action as vectors, nearest first) and a word index (the exact words). Their two lists are fused, a measured trust nudge is added (checked beats unchecked, official beats unofficial, used beats unused, demos drop), the top twenty are re-read by a second model that scores each against the question, and trust is applied once more at the end. Nothing in that chain can change a label.