People Search · Composio Instant

Describe the hire in plain words ("staff ML engineers in SF who've shipped LLM inference") and People Search's AI sourcing agent turns the brief into a ranked shortlist through Composio Instant. It finds candidates with Apollo and other data providers, has Jev, TypeSafe AI's evaluation model, score each one against your must-haves, and writes a note on why each person fits. Every call is billed to the signed-in user's Composio wallet with no API keys or subscriptions, and you see the exact price of each email before you pay for it.

• Sign in is the OAuth flow of connect.composio.dev/mcp (OAuth 2.1, PKCE, dynamic client registration). The bearer token it returns is what the agent sends to the Composio MCP server, so every Instant call bills the user who signed in.
• Search is free. Apollo people search returns titles, companies, first names and masked surnames at $0.
• Jev scores the shortlist. Every candidate is scored against the brief's must-haves by Jev, TypeSafe AI's evaluation model on Composio Instant, at a tiny fraction of a cent per candidate, and the board is ranked by that score.
• Reveal costs cents. Each person on a shortlist has a Reveal button with its price. Revealing tries providers cheapest-first and stops at the first email: Apollo enrichment unmasks an Apollo surname, then Tomba and Hunter, which bill only when they find one. Bulk reveal, email verification, a saved list and CSV export are on the board.
• Business development. Ask which companies are hiring for a role and the agent searches LinkedIn, Indeed, Glassdoor and Google Jobs and shows the postings as cards. "Who's hiring" on any card finds the likely hiring manager and in-house recruiter.
• Intros. Ask for a warm intro to a candidate or hiring manager and the agent finds the person, researches recent hooks (LinkedIn and X posts via ScrapeCreators, podcast appearances via Listen Notes, company news via Akta and Context.dev, hiring signals via PredictLeads, papers via OpenAlex), then writes a forwardable intro as a fenced ```intro block in its reply. The block streams token by token and renders live as an email card with sourced hooks, Copy and Open in Gmail.
• Every call is visible. The run trace shows each tool call with its provider, arguments, price, duration and raw response, and the header meter tracks spend for the session.

Run it

pnpm install
cp .env.example .env.local   # then fill it in, see below
pnpm dev

Open the URL next dev prints and sign in with Composio. New Composio wallets start with $2 of credit.

Environment

Variable
OPENROUTER_API_KEY	Required. The agent runs on AGENT_MODEL through OpenRouter (AI SDK OpenRouter provider).
AGENT_MODEL	Defaults to google/gemini-4-argon ($2 / $10 per M tokens). deepseek/deepseek-v4.1-flash ($0.14 / $0.42), minimax/minimax-m3.1-flash, z-ai/glm-5.3-flash and anthropic/claude-opus-5.5 also work.
SESSION_SECRET	32+ random characters; encrypts the session cookie (openssl rand -base64 48).
APP_URL	Optional. Fixes the OAuth redirect origin when running behind a proxy.

The OAuth client is registered automatically for whatever origin the app runs on, so a different port or domain needs no setup.

How it works

The chat follows the Composio playground (ComposioHQ/dashboard, src/app/api/playground and src/components/core/chat): Vercel AI SDK v6 streamText inside createUIMessageStream, sent as the SSE UI message stream, read with useChat (50ms throttle), rendered in time order with "Working · 12s" groups that stay open while they run and fold away after, tool chips with live timers, a live reasoning preview, and Streamdown for text. Fast models deliver whole paragraphs in one burst, so the server paces text and reasoning out word by word. Tool outputs are compacted before they reach the browser; full conversations stay on the server.

Reloading picks up where you left off. Each signed-in user's chats (full tool outputs included) and workspace (open chat, revealed contacts, saved list, spend total, spend cap) are saved as JSON under .data/ (gitignored) by src/lib/persist.ts, and the page renders them on load. That's a local store for a demo; on a serverless host, swap it for a database.

Costs are shown at the precision Instant bills in ($0.0015, not $0.002). A Reveal button shows the full range across every provider it may try, from the first one finding the email to the worst case, with the breakdown on hover. The per-request spend cap checks the top of each price range times the number of results a call asks for, so a batch can't exceed it. "Spent" adds up the charges Composio actually reports.

browser ──/api/auth/login──▶ connect.composio.dev/oauth/authorize ──▶ /api/auth/callback
   │                                                                      (encrypted httpOnly session)
   ├──/api/chat (AI SDK UI message stream) ──▶ streamText ──▶ model via OpenRouter
   │                                    └──────▶ connect.composio.dev/mcp (Bearer = user token)
   │                                               COMPOSIO_SEARCH_TOOLS / MULTI_EXECUTE_TOOL
   └──/api/reveal ──▶ connect.composio.dev/mcp  (Reveal buttons skip the model entirely)

• src/lib/agent.ts: streamText with three Composio meta tools (wrapped so they never throw and report durationMs, charge and found people) Boards are fenced blocks in the reply (```people, ```companies, and ```intro for drafts), so they stream row by row like the text around them; src/lib/blocks.ts parses them, half-written or not. Enforces the per-request spend cap and a research budget, refuses workbench calls, and stops and retries a generation that writes a tool call as text or runs away. The MCP tool list is cached per sign-in, so a follow-up starts the model while the connection opens. Logs per-phase timings as [agent] lines.
• src/app/api/reveal + src/lib/reveal-*: Reveal tries Instant providers one at a time, cheapest first (Tomba, then Hunter, then Apollo; Apollo first when the surname is masked, since only Apollo can resolve its own record; ZeroBounce to verify), stops at the first email, and streams each attempt and its charge to the row.
• src/lib/entities.ts: turns any provider payload into people and companies (Apollo ids, masked surnames, Hunter verification, ContactOut profiles keyed by LinkedIn URL).
• src/components/: the workspace. chat-state.ts derives contacts and spend from the useChat messages and lays each message out into groups, text, boards and notices; timeline.tsx and text-segment.tsx render them.

Design follows brand.composio.dev: Geist and JetBrains Mono only, sans at 400/500, achromatic surfaces with #51a2ff as a spot colour for interactive state, and the Instant gradient for hero moments. Light and dark themes both work.

License

MIT. See LICENSE.