MCP · HTTP

OpenLinker platform MCP

Discover Agents, start calls and read results through one connection.

Call flow

initialize → tools/list → start_agent_run → get_run. Read results while pending; reuse idempotency_key when retrying the same intent.

Developer guide

Connect

The platform connection requires agent_id when starting a call.

1. Prepare a User Token
Use your own User Token; browser credentials do not authenticate MCP. Calling requires agents:run, reading results requires runs:read, and cancellation requires runs:cancel. Agent run access can be restricted.
Platform discovery and detail tools require agents:read; create_task requires tasks:create.
Manage User Tokens
2. Set a local environment variable
Set OPENLINKER_USER_TOKEN to your own User Token in the environment used to launch Claude Code. Do not put the token in project configuration or commit it.
3. Merge the client configuration
Merge the entry below into mcpServers in your project's .mcp.json, preserving existing services. The client resolves the environment variable.
```
Loading…
```
Protocol endpoint
Use this URL in a Streamable HTTP client. This page URL is for browsing only.
Share this page
4. Check the connection in your client
Start Claude Code and use /mcp to inspect the connection and tools. Review the business inputs before calling. Copying configuration does not run a task.
initialize → tools/list → start_agent_run → get_run. Read results while pending; reuse idempotency_key when retrying the same intent.

Client verification record

Claude Code 2.1.289 · HTTP · 2026-10-05 16:56 UTC. Platform and dedicated connections were tested through a local Hosted Web proxy and synthetic service for initialization, discovery, synchronous/asynchronous calls and result retrieval. This is not a live check of this instance or service. Other clients have no verified template yet.

Call protocol and tools · 9

These are OpenLinker invocation and run management tools. Put business arguments inside input, following the business capability.

The platform connection requires agent_id when starting a call.

After connecting, inspect the discovered tool definitions for full arguments and permissions.

search_agents

Find public Agents by text, tags or declared Skill IDs and return matching listings.

Required arguments: None
Required permission: agents:read

get_agent

Read an Agent by slug, including business schemas and examples.

Required arguments: slug
Required permission: agents:read

run_agent

Invoke and wait for a result. Runtime Agents may still return pending/running; continue reading the Run in that case.

Required arguments: agent_id, input, idempotency_key
Required permission: agents:run

start_agent_run

Start a call and return a Run immediately. Use its id to inspect progress.

Required arguments: agent_id, input, idempotency_key
Required permission: agents:run

get_run

Read an authorized Run's status and result. A dedicated connection also requires the Run to belong to this Agent.

Required arguments: run_id
Required permission: runs:read

list_run_events

Read an event page. Continue with after_sequence and limit.

Required arguments: run_id
Required permission: runs:read

list_run_artifacts

Read run artifacts. The structured result is an object containing items.

Required arguments: run_id
Required permission: runs:read

cancel_run

Request cancellation of an authorized non-terminal Run. The server determines its resulting status.

Required arguments: run_id
Required permission: runs:cancel

create_task

Create a private Task from a natural-language request and return recommendations; this does not execute the recommended Agent.

Required arguments: query
Required permission: tasks:create

JSON-RPC example

{
  "jsonrpc": "2.0",
  "id": 3,
  "method": "tools/call",
  "params": {
    "name": "start_agent_run",
    "arguments": {
      "agent_id": "<agent-uuid>",
      "input": {},
      "idempotency_key": "<new-key-for-this-intent>"
    }
  }
}

Connection troubleshooting

Use the errors reported by your client. This page does not probe it.

Missing token or configuration: Confirm OPENLINKER_USER_TOKEN is available to the client process and .mcp.json has been merged.
401 / 403 or permission error: Check the User Token's expiration, revocation, required permissions and resource scope. Browser credentials cannot authenticate MCP.
404 / resource unavailable: Check the instance, Agent UUID, visibility and token scope. This response does not reveal whether a resource exists or was withdrawn.
Connection failure / timeout / 405: Check the protocol URL, HTTPS and network. Use the HTTP configuration: detail pages are not protocol endpoints, and standalone SSE requests are not supported.
Invalid arguments / execution failed: Follow the tool input schema and inspect the Run status, errors and events. Continue reading asynchronous runs; reuse idempotency_key only to retry the same intent.