Documentation

Automation

Connect an Agent

Configure any MCP or API client to use Coil safely across reads, purchasing, operations, reporting, and audits.

Start With the Agent's Capabilities

Coil does not run the AI model. Your agent uses Coil's tools/API with a workspace-scoped key. MCP clients can discover tools; HTTP-capable agents can use the documented API directly. A public docs URL requires no credentials, but execution always does.

  1. Identify the intended tasks: read-only analysis, receipt/purchase processing, sales report imports, stock operations, or bookkeeping.
  2. Check whether the client supports remote Streamable HTTP MCP and a fixed Authorization header. An OAuth-only client is not supported by the current connector.
  3. For binary uploads only, confirm access to the original file and a shell/binary HTTP client. If those are unavailable, use Coil's manual upload workflow.
  4. Check network policy before promising a connection. The MCP endpoint and production upload host are on www.usecoil.app; no upload redirects are used.

Create a Workspace Key

  1. A workspace owner opens Settings > Agent Access and creates a named, expiring key.
  2. Select the smallest preset/permissions that cover the intended tasks. Add finance, deletion, or other privileged scopes only deliberately.
  3. Enter the key in connector settings or a local secret manager. Never paste it into the conversation or documentation URL.
  4. Use separate keys for separate agents/tasks, and revoke a key when it is no longer needed.

Configure Remote MCP

The JSON is a generic URL-capable client's example, not a promise that every client accepts the same configuration file. Use your client's connector UI or current configuration format. Set network allowlists separately where your organization requires them.

  1. Add a remote/custom connector named Coil. In current Claude clients this is under Customize or Settings > Connectors.
  2. Set the server URL to https://www.usecoil.app/api/mcp.
  3. Choose the fixed-credential/no-OAuth option, currently labeled No sign in in Claude, and add Authorization: Bearer YOUR_COIL_API_KEY as a request header.
  4. Enable the connector for the conversation and review tool-approval settings. Do not automatically approve financial or destructive actions.
  5. Reconnect after API changes so the client refreshes schemas. tools/list exposes only tools authorized by this key, not every tool in the public reference.
json
{
  "mcpServers": {
    "coil": {
      "url": "https://www.usecoil.app/api/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_COIL_API_KEY"
      }
    }
  }
}

Configure Direct API Access

All operation calls use POST /api/automation/OPERATION with a bearer key and JSON. GET /api/automation/schema is authenticated, exposing only contracts available to that key. /docs/api.json is the public, full reference. Existing UI-session routes are not unlocked by API keys.

bash
curl "https://www.usecoil.app/api/automation/list_products" \
  -H "Authorization: Bearer $COIL_API_KEY" \
  -H 'Content-Type: application/json' \
  --data '{"limit":25}'

Verify a Read Before a Write

Start with list_products or get_inventory_summary. Confirm the response is from the intended workspace and that expected tools are discoverable. Then read the relevant record/version or prepare a preview before requesting changes.

text
Read https://www.usecoil.app/docs/agents and the linked reference.
Tell me what your current client supports, the permissions my tasks require,
and the configuration I must enter locally. Do not ask me to paste a key in chat.
Start with read-only checks; show proposed changes before applying them.

Delegation Boundaries

No agent tool can manage team access, keys, billing, global administrators, or workspace time zone and default tax settings. Giveaways use a dedicated reviewed preview and permission; other stock-removal reasons are not exposed. Do not promise an unsupported task just because a similar UI operation exists. Discover the actual permitted tools and use the workflow guide for reads, versioned edits, previews, and retries.