Operis

Operis

🎬 Our hackathon demo

This is the demo video we submitted to the Naïve × YC Autonomous Companies Hackathon (2:00).

▶️ Watch the demo (MP4)

What it shows: a Naïve-run company needs something only a person can do, so its agent makes one call to Operis with the job, the budget and what counts as done. Operis finds a real person, invites them, and they accept the shift and do the job in the real world. Their proof is checked and they get paid. No human at Operis touches any of it.

📑 Presentation slides

The slides we presented with (15 slides): the team, the problem, how one call becomes a verified, paid human action, the opportunity and our vision.

📄 Open the slides (PDF)

🗺️ New here? Start with docs/FLOW.md: the full job flow, who owns each step, and which folder your edits go in.

🤖 Want your agent to hire a human? Get a key at https://operis.live/developers, or read docs/CONNECT.md (live at https://operis.live/llms.txt).

🧪 Testing? docs/TEST_PLAN.md · 🚀 Running it in production? docs/GO_LIVE.md

Live

URL	What
https://operis.live	Website (Dora's app): home, worker job pages /j/<token>, /jobs, /dashboard
https://operis.live/developers	Agent companies: self-serve API key, connect instructions, copy-paste agent prompt, key test
https://admin.operis.live	Founders: live requests by company, agent activity, proofs (sign in with your founder email)
https://operis.live/mcp	Public MCP server (request_human, get_job, cancel_job), proxied to the API
https://operis.live/llms.txt	Guide for agents
https://api.operis.live	API directly (MCP, REST, webhooks); founder console at /admin

Give this to your agent: set up https://operis.live/SKILL.md (it signs you up, saves its key, and connects).

Connect any AI app with no key: add https://operis.live/mcp as a connector / MCP server (Claude.ai, ChatGPT, Claude Code, Cursor, VS Code) and sign in when asked (OAuth).

Use Operis from Claude Code (MCP server + "hire-a-human" skill), with OPERIS_API_KEY from https://operis.live/developers in your environment:

/plugin install operis --marketplace cubic1M/Naive-Autonomous-Company-Hackathon-Project

Humans on demand for autonomous companies. An AI-run company asks for a human through our MCP server. Operis finds the right people online, reaches them through contact info they already publish, and pays whoever finishes the job. Workers never have to sign up in advance.

Built at the Naïve × YC Autonomous Companies Hackathon. Operis is itself an autonomous company: six Naïve agents (CEO, Ops, Verifier, Finance, Sales, Disputes) run it in the Operis Naïve org, and the founders only set limits and approve big decisions.

---

How it works

Autonomous company ──MCP──▶ Operis API ──▶ Naïve agents (CEO, Ops, Verifier, Finance)
                                                 │
                                   1. find people near the job (web scraping)
                                   2. message them via their public contact
                                                 ▼
                              Human clicks link ──▶ Operis job page
                                                 │  Accept / Decline
                                                 │  Upload proof (photo / video / file)
                                                 ▼
                              Verifier checks proof ──▶ Customer charged ──▶ Worker paid

1. Request. A customer agent calls request_human on our MCP server with the task, location, deadline, budget, any certification needed, and what proof it wants.
2. Safety check. We reject illegal, identity-faking, surveillance or licensed work. Borderline requests go to a founder.
3. Hold payment. The customer's card is held, not charged (Stripe, manual capture).
4. Find people. The scraper looks for people near the job on public, work-related sources: gig boards, campus job boards, business listings, and past workers.
5. Reach out. We message many people at once using contact info they published themselves (email, SMS, etc.), with a link to the job page. The first person to accept gets the job. Every message includes an opt-out.
6. Job page. The human sees the job description, pay, location and required certification, and clicks Yes or No. If needed, they upload extra info such as a certification.
7. Proof. When they're done, they upload the proof the job asked for: photos, video or documents, with GPS and a timestamp.
8. Verify. The Verifier agent checks the proof against the request. It approves if it's confident; otherwise a human reviewer decides.
9. Pay. The customer's card is charged. The worker connects a bank account once, on their first payout, and receives their share. We keep the fee.

Money

• Workers pay nothing. No signup, no fees.
• Customers pay only for finished jobs. The card is held up front and charged only when proof passes. If the job fails, the hold is released.
• Our cut: 25% of the job price ($20 minimum job). Example: a $60 job pays the worker $45, and Operis keeps $15.

MCP tools (for customer agents)

Tool	What it does
request_human	Post a job: task, location, deadline, budget, certification, proof requirements
get_job	Status, plus the proof package once approved
cancel_job	Cancel before someone accepts

Example request:

{
  "title": "Photograph the Ferry Building entrance and posted hours",
  "instructions": "Stand in front of the main entrance on The Embarcadero. Take one photo of the entrance and one close-up of the posted hours.",
  "location": { "address": "1 Ferry Building, San Francisco, CA 94111", "lat": 37.7955, "lng": -122.3935 },
  "deadline": "2026-11-08T14:00:00-08:00",
  "budget_usd": 60,
  "certification_required": null,
  "proof": { "photos": { "min": 2, "must_show": ["main entrance", "posted hours"] } }
}

The agent team (Naïve)

Agent	Job
CEO	Owns the board, assigns every job, P&L, disputes, daily report to founders
Ops	Runs the scraper, sends outreach, picks the worker who accepts
Verifier	Checks proof (vision model + GPS); approves or escalates
Finance	Holds, charges, payouts, refunds, ledger
Sales	Gets other teams' agents connected to our MCP server

Limits set by the founders: a daily spend cap per agent; refunds and payouts over $150 need the CEO agent's approval; anything over $1,000, or any legal or safety issue, goes to a founder; jobs only within the demo area (San Francisco).

Team split

Role	Owner	Builds
Frontend	Dora	The worker website: job page (description, pay, certification, Yes/No), upload forms, proof upload with GPS, payout setup screen, plus the live dashboard (map, live job, agent feed, money)
Scraping & outreach	Andrew	Finding people near a location, extracting public contact info, sending the "blast" (email / SMS), first-yes-wins logic, the regulars list
Verification & payouts	Cameron	Proof checks (GPS distance, timestamp, image/video review with a vision model), approve/escalate flow, Stripe hold → charge → payout to bank
Naïve agents & MCP	Edwin	The MCP server and API, database, the five Naïve agents (prompts, budgets, tool permissions, team board), and wiring everything together

Shared contract: everyone builds against the same jobs / workers / proofs tables and the job states below. Change them only after agreeing as a team.

requested → held → outreach → accepted → proof_submitted → approved → paid
                                                         ↘ escalated / rejected
(also: cancelled, expired, refunded)

Stack

Layer	Tool
Agents	Naïve Managed Agents (@usenaive-sdk/managed-agents, naive CLI)
API + MCP server	TypeScript on Vercel (@modelcontextprotocol/sdk)
Database	Postgres (Neon, or a Naïve fullstack app)
Payments + payouts	Stripe Checkout (manual capture) + Stripe Connect Express
Messaging	Twilio SMS + email via Naïve inboxes
Proof storage	Vercel Blob (or Naïve app storage)
Vision	A vision model via Naïve's model router

Rules we follow

• Only public, work-related contact info. We don't buy or dig up personal data. Everyone we message can opt out, and once they do, we never contact them again.
• No identity work. Humans never sign contracts, pass CAPTCHAs, receive OTPs or pretend to be the agent.
• No licensed advice sold as ours. If a job needs a certification, the worker shows it, and they do the work in their own name.
• Every job is verified. GPS, a timestamp and proof are required before money moves.

Getting started

npm install                                   # repo root: npm workspaces (apps/*, services/outreach, services/money-verification)
docker run -d --name operis-pg -e POSTGRES_PASSWORD=dev -e POSTGRES_DB=operis -p 55433:5432 postgres:17
cp apps/api/.env.example apps/api/.env.local  # fill in secrets (see apps/api/README.md)
npm run db:migrate                            # all migrations (shared schema, API, outreach, money-verification)
npm run dev                                   # API on http://localhost:3001
cd apps/web && cp .env.example .env.local && npx next dev -p 3100   # website on http://localhost:3100 (proxies /api → :3001)
npm run smoke                                 # 20 end-to-end checks against the API

Useful scripts: npm run approvals (founder approvals for Naïve agents), npm run agents:setup (apply agents/agents.json to Naïve), npm run simulate -- <job_id> (play the worker), npm run customer:create.

Full flow, ownership and build order: docs/FLOW.md. Shared contracts: contracts/. Production setup: docs/GO_LIVE.md.