Agent Seek — Less SEO. More signal.

Cheap web recall (You.com) + Jev taste → Google-simple UI, HTTP API, and an agent Skill.

Google-simple for people. Judgment-native for agents.

You.com discovers web candidates. TypeSafe Jev cascade-ranks them. Clients get a small scored JSON list (plus a simple search UI). Agent Seek does not write the answer. Web results are scored for prompt injection before the agent reads them (signals.prompt_injection; UI: Injection risk). If hard gates empty the scored list, restored rows set gates_relaxed: true.

• Live demo: https://agentseek.dev

• Not Deep Research / Deep Agents. Max 100 discover candidates per query.

• API/MCP default mode=snip (title/URL/snippet; cheaper/faster). Pass mode=deep for Stage A survivor fetch (cap 12). Website UI is fixed snip.

Two ways to try it

1. Live demo (no clone): paste the public AS_… Bearer into MCP or curl — see Try the live demo. This burns operator You.com / TypeSafe quota and the per-IP demo limit on agentseek.dev.

2. Self-host / local: clone + .env — you must supply your own YDC_API_KEY and TYPESAFE_API_KEY. Local AGENT_SEEK_API_KEY is only your machine Bearer (not the live AS_… key).

Related MCP server: agent-web-search-mcp

Try the live demo (easiest path)

The live host publishes a public demo key (not a secret). It is the same AS_… value the website injects via /config.js. Anyone can copy it. It only spends operator upstream quota / the per-IP demo limit on agentseek.dev — it does not unlock You.com or TypeSafe secrets.

AS_LXxYWY1wQcEwliYnkfAXRA-c0sPlsRtVz4j_ZvJRlnI

Cursor MCP (live)

Paste into your Cursor mcp.json:

{
  "mcpServers": {
    "agent-seek": {
      "url": "https://agentseek.dev/mcp",
      "headers": {
        "Authorization": "Bearer AS_LXxYWY1wQcEwliYnkfAXRA-c0sPlsRtVz4j_ZvJRlnI"
      }
    }
  }
}

Curl (live)

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,"max_candidates":50}'

Zero-auth inspect (no key): GET https://agentseek.dev/v1/sandbox, GET /health, GET /openapi.json.

search_web accepts Bearer (public demo key on live, or your local AGENT_SEEK_API_KEY) or OAuth scope search:read. Health/docs helper tools are unauthenticated.

Quickstart — local clone

Python 3.12+. Clone this repo:

git clone https://github.com/Gitmaxd/agent-seek.git
cd agent-seek

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

cp .env.example .env

Fill .env (never commit it):

Self-hosters: real search needs paid You.com + TypeSafe keys; without them only tests, fixtures, and the sandbox work. Tests run without live keys (pytest -q uses fixtures). Local POST /v1/search / MCP search_web needs those two paid upstream keys plus a local AGENT_SEEK_API_KEY.

uvicorn apps.api.main:app --host 0.0.0.0 --port 8787

Open http://127.0.0.1:8787 — the UI calls same-origin /v1/search with the injected Agent Seek key (never exposes You.com / TypeSafe keys). The website UI is fixed snip (no Mode control). API/MCP default is snip; pass mode=deep for Stage A survivor fetch.

Curl (local)

source .venv/bin/activate
export AGENT_SEEK_API_KEY="$(grep ^AGENT_SEEK_API_KEY= .env | cut -d= -f2-)"

curl -sS http://127.0.0.1:8787/health

curl -sS -X POST http://127.0.0.1:8787/v1/search \
  -H "Authorization: Bearer $AGENT_SEEK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"q":"Introducing System One Models Jev","k":10,"max_candidates":50}'

# Opt-in deep: Stage A survivor fetch (cap 12; slower / more expensive)
curl -sS -X POST http://127.0.0.1:8787/v1/search \
  -H "Authorization: Bearer $AGENT_SEEK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"q":"when were langchain deep agents released","k":10,"max_candidates":50,"mode":"deep"}'

Local MCP

After cp .env.example .env, set AGENT_SEEK_API_KEY=dev-agent-seek-key-change-me (or any local secret). That key is only for your machine — it is not the live AS_… demo key.

{
  "mcpServers": {
    "agent-seek": {
      "url": "http://127.0.0.1:8787/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_LOCAL_AGENT_SEEK_API_KEY"
      }
    }
  }
}

Optional / advanced: OAuth DCR

Agents that prefer Dynamic Client Registration can use POST /oauth2/register and the flows in /auth.md. You do not need OAuth to try the live demo — copy the public Bearer key above. Full tool contract: skills/agent-seek/SKILL.md.

Tests

pytest -q

No live API keys required. Optional live evals (soft-fail):

# Legacy substring hit-rate (evals/queries.json)
AGENT_SEEK_LIVE=1 python scripts/eval_compare.py

# Pairwise LLM-judge suite — dry-run needs no keys
python scripts/eval_llm_judge.py --dry-run

Frozen first run: evals/public_v1/published/v1.0.0-first-run/. Public page: /eval (markdown: /eval.md). Suite contract: evals/public_v1/README.md.

What it is

A small FastAPI service:

1. Discover — You.com web candidates (hard cap 100)

2. Rank — TypeSafe Jev cascade (subject / republisher / prompt-injection gates)

3. Snip (API/MCP default) — Stage B on title/URL/snippet. Deep (pass mode=deep) fetches main text for Stage A survivors (cap 12), then Stage B

4. Return — top-k scored URLs, snippets, flags, and signals

On total Jev failure: HTTP 200 with meta.ranking: "raw_fallback".

apps/api/          FastAPI (health, /v1/search, OAuth, MCP, serves UI)
apps/web/          Google-simple static UI + public agent/human prose
packages/core/     discover (You.com) + Jev cascade + cache + deep fetch
skills/agent-seek/ Agent Skill (well-known serves SKILL.md)
fixtures/          Recorded You.com + Jev payloads (keys redacted)
tests/             Offline pytest (no live upstream)
evals/queries.json Frozen legacy eval queries (optional live compare script)
evals/public_v1/   Pairwise LLM-judge suite + published first run (`/eval`)

MCP search_web does not accept REST-only nocache / rank. See /docs, /developers, /api/docs.

Progressive Disclosure Search Ranking

You.com discovers candidates. TypeSafe Jev (a System One model) re-scores them in stages. Agent Seek returns a small scored URL list; it does not write the answer. The funnel exists so you can start at the head of the list: prefer the top 1–2 ranked sources before expanding. One good keeper beats a context window full of searches.

Ranking is progressive: cheap judgment first, then fetch and deeper judgment only on survivors.

1. Discover — up to 100 You.com candidates (default 50 in)

2. Stage A — Jev scores coarse relevance across the batch; weak hits drop

3. Disclose — API/MCP default mode=snip skips fetch and ranks title / URL / snippet only (fast/cheap). Pass mode=deep to fetch main text for up to 12 Stage A survivors (not a full-web crawl; slower/more expensive). Website UI is fixed snip.

4. Stage B — Jev scores the richer evidence, then Agent Seek composes a 0–1 score, collapses near-dupes, applies hard gates, and returns top-k

Stage B is where page text (not just SERP blurbs) drives answerability, fact-stated, authority, and injection-risk judgments.

On total Jev failure: HTTP 200 with meta.ranking: "raw_fallback".

Jev metrics

Returned under signals when parsed:

Hard gates (order, 0.55): subject_match → is_republisher → prompt_injection. subject_match and is_republisher fail-open if parse missing; prompt_injection is fail-flagged (missing/unparseable dropped from results). Gated rows stay in raw_results. If hard gates empty the scored list, pre-gate ranking is restored and each restored row has gates_relaxed: true (relaxed-safety response; not a SearchMeta field).

Score composition favors answerability, with a smaller weight for stating the sought fact and a gated authority boost when scores are close — primary sources preferred without crushing a page that clearly answers.

A pairwise LLM-judge harness compares You.com order vs Agent Seek cascade on the same discover pool (evals/public_v1/). /eval shows snip preference beside the top-3 gold hit rate. The frozen deep run is in evals/public_v1/published/v1.0.0-first-run/.

Rate limiting (optional)

Local/dev is unlimited unless you set both UPSTASH_REDIS_REST_URL and UPSTASH_REDIS_REST_TOKEN. Missing either disables limiting — no Redis calls.

When enabled: sliding 60s window per client IP (not the shared UI API key), default AGENT_SEEK_RATE_LIMIT_PER_MIN=60, on search and MCP search_web only. Client IP is the rightmost public X-Forwarded-For hop, then X-Real-IP, then the socket address. Redis/network errors fail closed (503). Burst 429 is code=RATE_LIMITED and may include Retry-After.

The public host can also apply a lifetime per-IP demo quota, gated by AGENT_SEEK_DEMO_MODE (default disabled; live agentseek.dev sets enabled at deploy). When enabled: AGENT_SEEK_DEMO_SEARCH_LIMIT (default 5, Upstash INCR, no TTL). REST /v1/search, /api/v1/search (the HTML UI and any other REST client), and MCP search_web share that allowance. Exhausted searches from that IP return 429 application/problem+json with code=DEMO_EXHAUSTED and no Retry-After. Unauthenticated MCP helper tools do not consume it. See .env.example.

Secrets

Never commit YDC_API_KEY, TYPESAFE_API_KEY, local AGENT_SEEK_API_KEY, OPENAI_API_KEY, or Upstash tokens. .env.example has placeholders only. The live AS_… value in /config.js is a public demo key (documented on purpose). How we handle that vs self-host hardening: SECURITY.md.

Community

• Contributing — tests, product locks (You.com / Jev / 100-cap)

• Security policy — report vulns to 131803031+Gitmaxd@users.noreply.github.com

• Code of conduct

• Release notes: CHANGELOG.md

License

MIT. SPDX-License-Identifier: MIT

Author

Built by Git Maxd (@gitmaxd). Follow on X for more.