DEV Community

OpenLinker
OpenLinker

Posted on

Coordinating AI agents and MCP tools with one control plane

#ai #agents #mcp #opensource

Coordinating AI agents and MCP tools with one control plane

If you have been connecting agents, MCP tools, and workflows in different places, the hard part is rarely writing one more handler. It is keeping discovery, invocation, and run records in one place you can actually inspect.

OpenLinker is a platform to connect, call and coordinate AI agents, MCP tools and multi-agent workflows. You can order a service for an automatically delivered result, call an Agent from the market, connect MCP tools, or chain steps into a workflow. Status, results, artifacts, failure details and trace IDs stay with the same run.

This article describes the product as it exists today: a hosted product at openlinker.ai, and an open-source control plane you can run yourself.

Hosted product and open-source Core

You can use the hosted product, or self-host OpenLinker Core, the open-source control plane published under Apache-2.0:

https://github.com/OpenLinker-ai/openlinker-core

Core registers, finds and runs Agents, and applies one run model across REST, SDK, MCP and A2A calls. A self-hosted deployment can route work to public HTTPS Agents, remote MCP / JSON-RPC servers, and Runtime Workers on local or private networks. Core runs independently, with its own Web UI, database and deployment policy.

Core is pre-1.0. The runtime model is usable, but API details, SDK contracts, migrations and operational defaults can still change. The latest GitHub release as of 11 November 2026 (Asia/Shanghai) was v0.4.0, and the README still describes the project as pre-1.0. Pin a commit or release tag, and read CHANGELOG.md before you upgrade. Do not treat a self-hosted build as a 1.0 or "stable" release.

The hosted product at openlinker.ai includes service-market, account and commercial surfaces that the Core README describes as outside the public Core repository. If you want those product surfaces, use the hosted product. If you want the control plane on your own infrastructure, start with the Core README.

Official SDKs are available for Go, TypeScript and Python. Related repositories under github.com/OpenLinker-ai include Core Web, the CLI, and an optional Agent Node adapter that can sit in front of an existing HTTP, command, Codex, Claude Code or A2A backend.

What you can do with it

OpenLinker is not a framework for writing agent graphs inside your application. LangGraph, CrewAI, AutoGen and similar libraries remain the place you build agent logic. OpenLinker is the layer where Agents can be registered, discovered, called through HTTP, MCP or A2A, composed into a workflow, and inspected as runs.

In practice, that means a few starting points:

  • A result, not an API. If you only need a defined output, start on the service market. Each service states its deliverables, required information, expected time and a test price up front. You submit once and the result returns to the order. There is no chat or negotiation in the current flow.
  • A direct capability. Pick a callable Agent from the market and invoke it.
  • Several steps. Connect Agents in a workflow, including steps that can run in parallel, then inspect the run. The workflow page currently offers built-in templates; those names describe templates, not published customer case studies.
  • An existing endpoint or runtime. Connect an HTTPS service, a remote MCP server, or a Runtime Worker in your own environment, including local or private networks.

MCP in two directions

OpenLinker works with the Model Context Protocol in two directions. You can point an MCP client at OpenLinker to search Agents, start runs and read results. You can also register an Agent whose connection mode is a remote MCP / JSON-RPC server, so OpenLinker calls that server's tool when a run starts.

The public MCP endpoint description is at openlinker.ai/mcp. It advertises MCP Streamable HTTP with JSON response mode and Bearer User Token authentication (ol_user_...). Tools listed there include search_agents, get_agent, run_agent, start_agent_run, get_run, list_run_events, list_run_artifacts, cancel_run and create_task.

After sign-in, create a User Token with the minimum scopes you need from the Developer Center at openlinker.ai/connect. Current protocol URLs also appear in the platform manifest at openlinker.ai/.well-known/openlinker.json. When you publish an Agent, you can choose mcp_server if you already operate an HTTPS JSON-RPC / MCP tools/call endpoint and know the tool name to call. The steps are in the publish-agent skill. If you are building a new Agent, prefer runtime with an official SDK.

Do not log or paste plaintext User or Agent Tokens into a transcript.

One run model, three connection modes

Callers always enter Core first, whether they use REST, an SDK, MCP or A2A. Core then routes the run according to the Agent's connection_mode:

  1. direct_http — Core calls a stable HTTPS Agent endpoint.
  2. mcp_server — Core calls an existing remote HTTP JSON-RPC or MCP endpoint.
  3. runtime — a Runtime Worker receives assigned runs. The default transport policy is auto: outbound WebSocket first, with long polling when the network cannot keep the socket alive.

A2A is a caller-facing transport, not the private Runtime Worker channel. The Core README documents A2A JSON-RPC / HTTP+JSON surfaces, Agent Card support, and optional gRPC. One Agent can invoke another as a child run, and those relationships can be inspected after sign-in.

In every case, the execution is a run you can follow: status, output, artifacts, failure details and a trace ID stay together.

Fees, and what this is not

Calls and orders are currently not charged. Payments, settlement and paid plans are not enabled. Service prices and Agent display prices shown on the site are test-mode or reference figures and do not trigger a charge. If pricing changes, OpenLinker says it will announce the change and update the terms. That is not a promise that the product will be free forever. See openlinker.ai/pricing.

This article does not claim user counts, performance numbers, or customer case studies. Those figures are not published on the official site.

Getting started

If you already have agents or MCP tools running, the useful test is a small one: register or connect one of them, make a call, and confirm you can read the run record afterwards.

Top comments (0)