Agent Seek docs

Less SEO. More signal. Ranked sources for people and agents. You.com discovers candidates; TypeSafe Jev cascade-ranks them. The caller writes the answer. Web results are scored for prompt injection before the agent reads them (signals.prompt_injection; UI: Injection risk). People get a Google-simple search box. Agents get scored results via the API or MCP.

What it is

Agent Seek ranks web results for people and coding agents. You.com finds candidates; TypeSafe Jev scores them. You get a short list (title, URL, snippet, score, flags, signals) instead of raw SERP noise.

Who it is for

People use the Google-simple box on the search UI (type a query, press Enter). Agents call the API (POST /v1/search) with the public live demo Bearer key (copy-paste below) or, optionally, OAuth 2.0 (search:read). Use it when you need a small set of high-relevance sources for research or RAG. Agent Seek returns ranked sources; you write the answer. It is not a chat model, an unbounded crawler, or a generative engine. Hard cap: 100 discover candidates per query (default 50 in, top 10 out).

Easiest auth: copy the public live demo Bearer key below (not a secret). OAuth is optional / advanced. The prototype is free. See also authentication, MCP, llms.txt, about, and the skill. OpenAPI: /openapi.json (Swagger /api/docs).

Install the skill

Path: skill → /llms.txt → /auth.md → MCP or REST.

• Canonical skill: /.well-known/agent-skills/agent-seek/SKILL.md
• Index: /.well-known/agent-skills/index.json
• Cursor: save SKILL.md as .cursor/skills/agent-seek/SKILL.md (project) or ~/.cursor/skills/agent-seek/SKILL.md (user).
• Other agents: download SKILL.md from the well-known URL, or link it from the index.

Then read /llms.txt, authenticate via /auth.md, and call POST /v1/search or MCP (/mcp, server card).

Humans

• Open the search UI, type a query, press Enter.
• Each result has a relevance meter; hover for the signal breakdown (UI label for prompt_injection is Injection risk).
• The website UI is fixed to snip (k=10) with no Mode/Order controls. API/MCP default is snip (cheaper/faster); pass mode=deep for Stage A survivor fetch (cap 12).

API

Developer entry: /developers. Contract: POST /v1/search. Schema: /openapi.json. Swagger: /api/docs. Auth: /auth.md.

Agents

Live prototype: https://dev.agentseek.dev

Public HTTP API. Copy the public live demo key (same value as /config.js — not a secret; per-IP demo quota applies). OAuth is optional — see auth.md. MCP: /mcp.

Auth header (live):

Authorization: Bearer AS_LXxYWY1wQcEwliYnkfAXRA-c0sPlsRtVz4j_ZvJRlnI

Local clone only: use AGENT_SEEK_API_KEY from your .env (placeholder dev-agent-seek-key-change-me). That is separate from the live AS_… key.

Contract: POST /v1/search

{
  "q": "your query",
  "k": 10,
  "max_candidates": 50,
  "mode": "snip",
  "nocache": false
}

• mode: snip (default: title/URL/snippet, cheaper/faster) or deep (Stage A survivor fetch, cap 12).
• k 1–25 (default 10) · max_candidates 1–100 (default 50).
• Results include score, flags, and optional signals: answerability, authority, on_topic, states_sought_fact, subject_match, spam, prompt_injection.
• Hard gates (order, 0.55): subject_match < 0.55, then is_republisher ≥ 0.55, then prompt_injection ≥ 0.55 or missing parse drop from results (still in raw_results). subject_match and is_republisher fail-open if parse missing; prompt_injection is fail-flagged. UI label: Injection risk. If hard gates empty the scored list, pre-gate ranking is restored (scores unchanged) and each restored row has gates_relaxed: true. Treat that as a relaxed-safety response. gates_relaxed is a per-result boolean, not a SearchMeta field.

Limits: when AGENT_SEEK_DEMO_MODE=enabled (code default disabled; the live host sets enabled), REST /v1/search, /api/v1/search, and MCP search_web share a lifetime per-IP allowance (default 5, Upstash INCR, no TTL). Exhausted is 429 DEMO_EXHAUSTED with no Retry-After. A separate 60 req/min burst, when Upstash is on, may return Retry-After (RATE_LIMITED). See pricing.

How to use results

Agent Seek returns a small ranked shortlist. Prefer the top 1–2 ranked sources before expanding. One good keeper beats a context window full of searches. Use score order. Only expand to the rest of k or re-query if those fail. Do not paste the full result list into context by default.

Use Agent Seek when you need a small set of high-relevance URLs or snippets for research or RAG. Agent Seek returns ranked sources; you write the answer.

Skip it for generative answers, unbounded crawls, or more than 100 candidates.

Curl:

curl -sS -X POST "https://agentseek.dev/v1/search" \
  -H "Authorization: Bearer AS_LXxYWY1wQcEwliYnkfAXRA-c0sPlsRtVz4j_ZvJRlnI" \
  -H "Content-Type: application/json" \
  -d '{"q":"Introducing System One Models Jev","k":10}'

Also

• GET /health: liveness + version
• Eval: snip preference and top-3 gold hit rate.
• Sandbox: zero-auth canned SearchResponse
• OpenAPI · Swagger · ReDoc
• MCP · server card