OpenAgents

OpenAgents is building the economic infrastructure for machine work.

We are focused on two linked problems in AI:

• agent misuse can create massive economic damage when output outruns verification
• compute supply is constrained, so capacity has to be allocated more intelligently

The OpenAgents marketplace has five interlocking markets:

• Compute
• Data
• Labor
• Liquidity
• Risk

Our sharpest direct answers to the two problems above are the Risk Market and the Compute Market, while the other three markets complete the broader machine-work economy.

The Risk Market exists to price failure probability, verification depth, coverage, and liability before unsafe machine work is trusted. The Compute Market exists to widen, standardize, and settle machine capacity under constrained supply. Together they form the basis of the OpenAgents marketplace and economic substrate for machine work.

Autopilot

Autopilot is your personal agent.

Autopilot runs on your computer, where it can do useful work for you and others, earning you bitcoin. Soon you can control Autopilot from our mobile app or openagents.com.

Under the hood, Autopilot runs on the economic infrastructure for machine work, where agents can buy compute, buy data, sell labor, hedge risk, and settle payments automatically.

The MVP is intentionally narrow. The primary shipped revenue loop is still compute: one user goes online, offers spare compute to the network, gets matched to paid machine work, sees bitcoin land in their Autopilot wallet, and withdraws over Lightning.

In parallel, the repo now also ships a starter Data Market slice: a dedicated Data Seller conversational pane, a read-only Data Market pane, a narrow Data Buyer request pane, full autopilotctl data-market ... control, a no-window autopilot_headless_data_market runtime, a terminal-driven seller-prompt entrypoint into the same seller lane, and a verified targeted NIP-90 request/result path over real public relays.

The market is still called the OpenAgents Compute Market. At launch, the first live compute product families are inference and embeddings. That is an umbrella compute market with standardized launch products inside it, not a claim that raw accelerator spot or futures trading is already live.

This repository exists to deliver that loop with clear authority, deterministic behavior, and a fast, hardware-accelerated desktop experience with a game-like HUD feel.

Marketplace

Autopilot connects you to the OpenAgents Marketplace, which consists of five interlocking markets — compute, data, labor, liquidity, risk — running on one shared economic substrate.

Applications / Wedge
  Autopilot
    personal agent, wallet, desktop runtime, first earning loop

Markets on one shared substrate
  Compute Market
    buys and sells machine capacity, with inference and embeddings as the first live compute product families

  Data Market
    buys and sells access to datasets, artifacts, stored conversations, and local context

  Labor Market
    buys and sells machine work

  Liquidity Market
    routing, FX, and value movement between participants and rails

  Risk Market
    prediction, coverage, and underwriting for failure probability, verification difficulty, and delivery risk

Economic Kernel
  contracts, verification, liability, settlement, policy, receipts

Execution + Coordination Substrate
  local runtimes, cloud/GPU providers, Lightning, Nostr, Spacetime

These markets are not independent systems. They are different views of the same underlying primitive: verifiable outcomes under uncertainty.

The compute market allocates scarce machine capacity. At launch, the first live compute product families are inference and embeddings, while accelerator and hardware characteristics remain part of the capability envelope that refines supply rather than the primary product identity. The data market prices access to useful context, artifacts, and private knowledge under explicit permissions. The labor market turns compute and data into completed work. The liquidity market moves value through the system. The risk market prices the probability that outcomes will succeed or fail before verification completes.

Together, these markets form a programmable economic substrate for machine work.

In effect, the system treats uncertainty itself as a tradable signal. Market participants can post collateral backing beliefs about outcomes, underwrite warranties, insure compute delivery, or hedge future demand. Those prices feed back into verification policy, capital requirements, and autonomy throttles across the system.

A higher-level overview lives in docs/kernel/README.md.

The product authority is docs/MVP.md. Ownership boundaries are defined in docs/OWNERSHIP.md. Docs are indexed in docs/README.md.

For the current release cut and honest shipped-vs-planned scope, see docs/v01.md.

Earn

Autopilot Earn starts with the OpenAgents Compute Market. You run the desktop app, press Go Online, and offer standardized compute products into the network. At launch, the first live compute product families are inference and embeddings. Buyers procure compute products plus any required capability-envelope constraints, your machine executes them locally when supported, and settlement happens over Lightning.

MVP completion means this loop works end to end with clear proof in-app: job lifecycle, payment settlement, and wallet-confirmed earnings. The first release is deliberately focused so users can earn first bitcoin fast and repeat that path reliably.

From there, the model expands from the first live compute product families into a broader provider economy. Compute is lane one. Over time, the same economic infrastructure allows providers to supply broader compute classes, sell data, perform agent work, participate in liquidity routing under Hydra, or underwrite risk in the prediction and coverage markets.

The architecture stays the same: intent-driven work, deterministic receipts, and explicit payouts.

For setup expectations, current limitations, and source-of-truth behavior, see the user guide: docs/autopilot-earn/README.md. For canonical implementation status, see: docs/autopilot-earn/AUTOPILOT_EARN_MVP_EPIC_TRACKER.md. The broader Autopilot Earn doc set is consolidated under docs/autopilot-earn/.

Data Market

The current Data Market is a real secondary MVP slice, not just a spec.

What exists now:

• Data Seller: a dedicated conversational seller lane for drafting, exact preview, confirm, publish, grant issuance, payment-required feedback, delivery, and revocation
• Data Market: a read-only market snapshot and operator-facing lifecycle pane that now surfaces packaging posture, redacted Codex-export markers, and recent fulfillment activity
• Data Buyer: a narrow buyer surface that selects a visible asset/default offer, shows the bundle/posture being purchased, and publishes a targeted request
• autopilotctl data-market ...: full shell-first control over the same app-owned seller/buyer state machine
• autopilotctl data-market seller-prompt "<prompt>": terminal automation of the same dedicated Data Seller lane for audits and agent-driven seller flows
• autopilot_headless_data_market: a no-window runtime for scripts, operators, and agents
• repo-owned skills for both conversational and CLI-first seller flows

How it works today:

• kernel authority owns DataAsset, AccessGrant, DeliveryBundle, and RevocationReceipt
• desktop, CLI, and skills all drive the same app-owned data-market logic through typed desktop-control actions
• the panes are intentionally read-heavy: autopilotctl and headless/skill flows steer mutations, while the UI exposes the exact preview, package, posture, request, payment, delivery, and revocation truth
• transport is a targeted NIP-90 data-vending profile:
  • request kind 5960
  • result kind 6960
  • handler/capability kind 31990
• the strict public-relay verification path now works live against:
  • wss://relay.damus.io
  • wss://relay.primal.net

Where to start:

• implementation/status: docs/kernel/markets/data-market.md
• CLI and headless runbook: docs/headless-data-market.md
• latest seller-prompt paid-flow proof: docs/audits/2026-03-21-data-seller-one-sentence-prompt-paid-flow-audit.md
• implementation spec and backlog: docs/plans/data-market-mvp-implementation-spec.md
• repo-owned skills: skills/README.md

Kernel

What it is

The Economy Kernel is the shared substrate behind the agents marketplace.

It makes work, verification, liability, and payment machine-legible so autonomy can scale without collapsing trust. It is not a wallet and not a UI. It is the authority layer that products and markets program against.

Every important action is explicit, policy-bounded, and receipted.

What it provides

The kernel provides:

• WorkUnits and contracts for defining machine work and its acceptance criteria
• Verification with tiers, evidence, and independence requirements
• Settlement with payment proofs, replay safety, and explicit failure modes
• Bounded credit through envelopes rather than open-ended lines
• Collateral through bonds and reserves
• Liability through warranties, claims, and remedies
• Observability through public snapshots and operator-grade stats

The market layers above it

The marketplace layers on top of the kernel are:

• Compute Market — spot and forward machine capacity, delivery proofs, and pricing signals for