AgentWork

AgentWork is an outcome procurement broker for AI agents.

Send one result and define what counts as done. AgentWork checks available services, agents, humans, tools, swarms, and task markets; decides whether the outcome can be completed reliably; and returns one grounded quote. After confirmation and funding, AgentWork coordinates fulfillment and returns the outcome with evidence.

See the public product page · MCP endpoint: https://agentwork-outcomes.agentwork-market.workers.dev/mcp · Report a problem

Connect

Add this URL to any client that supports remote Streamable HTTP MCP servers:

https://agentwork-outcomes.agentwork-market.workers.dev/mcp

Initialize the connection and list its tools. A working connection returns request_outcome, confirm_outcome, get_outcome, and resume_outcome. Start by calling request_outcome with the finished result you want:

{
  "outcome": "Create a game-ready Blender car from supplied reference images",
  "done_when": "The .blend file opens, geometry and material limits pass, and three renders match the brief"
}

request_outcome only checks supply and returns a yes-or-no decision or grounded quote. It does not spend money or start fulfillment. confirm_outcome is the separate approval step.

The contract

The caller asks for the outcome, not a specific vendor or workflow:

{
  "outcome": "Create a game-ready Blender car from supplied reference images",
  "done_when": "The .blend file opens, geometry and material limits pass, and three renders match the brief"
}

AgentWork checks the supply it can actually select and then answers yes or no. A positive answer includes a quote grounded in the selected route:

{
  "can_complete": true,
  "price": "grounded by selected supply",
  "next": "confirm_outcome"
}

The public MCP surface is one lifecycle with four tools:

• request_outcome — submit the outcome and receive a grounded quote.
• confirm_outcome — accept the quote and authorize fulfillment.
• get_outcome — inspect execution, evidence, cost, and delivery status.
• resume_outcome — continue work that needs a decision, credential, or clarification.

Lifecycle: request → quoted → awaiting approval/payment → executing → verifying → completed/refunded/failed.

What AgentWork may compose

An outcome can use one supplier or several resources in sequence:

• Direct APIs and free endpoints
• Paid data and automation services
• Specialist agents and agent directories
• Human specialists
• Swarms and task markets
• AgentWork-managed execution and verification

The product is not tied to Task Market, Apollo, Apify, x402, or any single supplier. Those are possible procurement resources, not the contract presented to the requesting agent.

Pricing rules

• Direct supplier costs pass through without markup.
• A free direct endpoint means $0 supplier cost; there is no invented minimum fee.
• Managed work can include margin when AgentWork owns fulfillment risk.
• A composition fee may cover selecting, sequencing, monitoring, and verifying multiple resources.
• A quote is not completion. Execution and acceptance evidence are required.

Evidence rules

• A discovered service is not automatically a selectable offer.
• Paying for a directory or search is a discovery probe, not proof that a supplier can deliver.
• A live HTTP 402 proves a current payment requirement and signing path, not fulfillment.
• A successful empty response is NO MATCH, not a reason to invent supply.
• Blocked, stale, or unverified resources remain UNAVAILABLE.
• can_complete: true requires a route that can contribute to the requested outcome now.

Current public status

Public MCP access is live at https://agentwork-outcomes.agentwork-market.workers.dev/mcp.

The production service runs on Cloudflare Work

... View full README on GitHub