orangeforge.ai Français Download

documentation

Publish an agent

Choose the entry point that fits your audience without changing the agent contract, versions or traces.

Make agents work together#

MCP and A2A let your agents communicate with external systems. Agents in the same orangeforge.ai team do not need that façade: coordination is built into both the Worker and Mission runtimes.

  • call_agent invokes a Worker agent and waits for its response in the current run.
  • start_mission starts an asynchronous mission and immediately returns its identifier.
  • check_mission follows its status and retrieves its deliverables.

Every delegation is signed by the platform, so agent code never receives a team API key. The call chain keeps its root, parent and depth so every branch can be reconstructed in traces. It is limited to five levels and cycles are rejected. Every child remains metered and billed like an ordinary run.

The agent prompt and code decide when to delegate and how to react to a failure. The platform secures and traces coordination, but does not retry a step on their behalf.

Try the public agent#

The horaires-demo demo uses all four production access methods:

Minimal MCP configuration:

{
  "url": "https://mcp.agent-forge.cc/mcp/xp_40112278877c704b5ac8dac1",
  "headers": { "x-api-key": "<EXPOSURE_KEY>" }
}

Chat and the HTTPS demo are available immediately. MCP and A2A invocation require an exposure key: the public identifier in the URL is not a secret and cannot authorize a call on its own. The A2A card remains readable without a key to support discovery.

Choose a channel#

An activated agent keeps the same contract, versions and traces regardless of its entry point.

NeedAccessCredential
A conversation to share or embedPublished ChatAnonymous session, magic link or host-signed identity
A tool in a compatible AI clientMCPfex_… exposure key
A remote agent for an orchestratorA2Afex_… exposure key
A custom server integrationHTTPSfak_… team API key

Publish a Chat#

In forge ui, open the project and then Chats. Create a surface, choose its agents, access mode, budgets, retention and allowed origins, then publish it.

The administration provides a standalone URL:

https://chat.agent-forge.cc/c/<publicId>

and a widget:

<script
  src="https://chat.agent-forge.cc/widget.js"
  data-surface="<publicId>"
  async
></script>

An anonymous surface receives a pseudonym scoped to the browser. An authenticated surface uses a magic link or a short identity token signed by the host website backend. The browser never receives a team API key.

Expose through MCP and A2A#

In forge ui, open Exposure. Create an exposure, select contracted Worker agents, enable MCP and/or A2A, then publish it.

Create an fex_… exposure key. It is shown once, opens only that exposure and can be revoked immediately. It is not an API key for your account.

An agent that can write to an external service or handle a sensitive capability requires explicit administrator opt-in. MCP and A2A V1 do not relay an end-user confirmation.

Connect an MCP client#

The administration provides the endpoint and command for the exposure:

claude mcp add --transport http forge-<publicId> https://mcp.agent-forge.cc/mcp/<publicId> --header "x-api-key: fex_…"

The server publishes one tool per contracted agent. Parameters come from the versioned input contract; every call remains metered and traced like an ordinary run.

Connect an A2A client#

A2A V1 exposes one card per agent:

https://mcp.agent-forge.cc/a2a/<publicId>/<slug>/.well-known/agent-card.json

The JSON-RPC invocation endpoint is:

https://mcp.agent-forge.cc/a2a/<publicId>/<slug>

A2A 0.3 is synchronous. Long-running tasks, streaming over A2A and the A2A 1.0 interface are not part of this version.

Call the HTTPS API#

For a custom server integration, invoke the activated agent directly:

curl -N https://agents.agent-forge.cc/my-agent/invoke \
  -H "x-api-key: fak_…" \
  -H "content-type: application/json" \
  -d '{"query":"..."}'

The fak_… key opens the team's agents: keep it on the server. An fex_… key is deliberately rejected by this direct endpoint.

End-user identity and shared controls#

A Chat can carry end-user identity. MCP and A2A V1 authenticate the client with an exposure key but do not yet carry the identity of the person using that client.

Every access method keeps input contracts, active versions, metering, runs and traces. Disabling a surface or revoking its key cuts off access without changing the agent files.