For developers

The same gateway
your agents use

A REST API for orchestration and an MCP server for tools. Both run behind Aperture, so tier routing and the evidence ledger apply to your calls exactly as they do to ours.

REST + streamingMCP stdio & HTTPHermes MIT-licensed
Three surfaces

Pick the layer you need

REST API

Create agents, dispatch tasks, read results, pull ledger entries. JSON in, JSON out, server-sent events when you want tokens as they arrive.

api.wildfire.business/v1

MCP server

Every connector your agents can reach is exposed as an MCP tool. Point Claude Desktop, an IDE, or your own client at it and inherit the same scopes.

mcp.wildfire.business

Hermes

The agent framework underneath, MIT-licensed. Run it against your own models if you would rather not run it against ours. No lock-in by design.

github.com/wildfire/hermes
Getting started

Authenticate, then dispatch

Tokens are workspace-scoped and inherit the tier policy of the agent you act as. There is no way to call a model your policy forbids.

Authenticate
# workspace token from Settings → Developers
export WILDFIRE_TOKEN="wf_live_..."

curl https://api.wildfire.business/v1/agents \
  -H "Authorization: Bearer $WILDFIRE_TOKEN"
Dispatch a task
curl https://api.wildfire.business/v1/tasks \
  -H "Authorization: Bearer $WILDFIRE_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "agent": "sofie-agent",
    "tier": 2,
    "prompt": "Summarise the #meridian thread and flag blockers",
    "tools": ["slack.history", "hubspot.deals"]
  }'
Connect an MCP client
{
  "mcpServers": {
    "wildfire": {
      "url": "https://mcp.wildfire.business/sse",
      "headers": { "Authorization": "Bearer wf_live_..." }
    }
  }
}
Reference

Core endpoints

EndpointMethodWhat it does
/v1/agentsGET, POSTList workspace agents, or create one with a name, owner and memory scope.
/v1/tasksPOSTDispatch work. Returns a task id immediately; stream tokens with ?stream=true.
/v1/tasks/:idGETStatus, output, credits consumed, and the tools the agent actually called.
/v1/vault/searchPOSTRetrieval against your knowledge vault, respecting per-agent memory boundaries.
/v1/connectorsGETWhich connectors are authorised, by whom, and with what scopes.
/v1/ledgerGETAppend-only evidence entries. Filter by agent, actor, tier or date. Article 14 export.
Rate limits

600 requests per minute per workspace, 60 concurrent tasks. Streaming connections do not count against the request limit. Limits are soft — we will call you before we throttle you.

Want the full
reference?