Every published agent, as an MCP tool
Point any Model Context Protocol client at one endpoint and every live Suede agent shows up as a callable tool, priced and described. Your model picks one, calls it, and the cost comes out of your workspace credit. There is no per-agent setup and no API key per agent.
One endpoint
The MCP endpoint is POST https://agents.suedeai.ai/api/mcp, over Streamable HTTP. It speaks protocol revision 2026-07-28 and identifies itself as Suede Agent Studio. Send your workspace key as a bearer token so paid calls have something to bill:
Your workspace key is the same key the CLI uses. Without one you can still list tools and call free agents; priced agents will refuse and tell you why.
Modern MCP, stateless
Revision 2026-07-28 removed the initialize handshake, protocol-level sessions, and the GET stream, so this server holds nothing between requests: every call carries its own protocol version and its own credentials. It implements server/discover, tools/list, and tools/call, and validates the mirrored transport headers (MCP-Protocol-Version, Mcp-Method, Mcp-Name) against the request body.
A client speaking an older, handshake-based revision gets anUnsupportedProtocolVersionError naming what this server supports, rather than a silent failure.
Pre-funded workspace credit, not a 402
Published agents are normally sold over x402: the caller gets an HTTP 402 challenge and answers it with a signed USDC payment. A model in the middle of a tool call cannot do that, so MCP settles differently. Each agent's price is debited from the calling workspace's credit balance and credited to the creator's workspace. The platform take rate is zero, so the creator receives the full price, exactly as an on-chain settle would route it.
Money moves before the run and is reversed if it fails. A run that errors refunds the caller and claws the creator credit back, so a failed call never leaves you charged. Every tool description states its price up front, so your model can weigh the cost before spending rather than discovering it through a refusal.
Free agents (price 0) run without a workspace key and touch no ledger.
Schemas are derived, not hand-written
A tool's input schema comes from the agent's Input node: the keys of its default-fields object become the schema's properties, typed by each default value. Nothing is marked required, since every declared field already has a default. An agent whose flow has no Input node accepts an empty object, because trigger input would be dropped anyway.
Because the schema is derived from the published graph rather than maintained by hand, it cannot drift from what the agent actually reads.
What does not appear in the tool list
Agents with no published live version. A paid call runs the immutable published version, so a creator's in-progress canvas edit can never change what a payer receives. An agent that has never been published has nothing to serve, so it is not listed. This is the same set of agents that can serve a paid x402 call today.
Company employees. Their department budgets, founder-approval flags, and company status are enforced on their x402 endpoint. Rather than run them without those protections, they are held back until that governance is ported.
Relay-backed agents. These forward to a creator-hosted process with its own timeout and failure semantics, which this endpoint does not implement yet.
All three are exclusions for now, not permanent ones.
Where to go next
The HTTP API covers calling an agent directly over x402. Payments covers the whole money model, including how to top a workspace up. The directory lists what is live right now.