Agent Harnesses

What agent-harnesses MCP server does

The agent-harnesses MCP server exposes a curated reference list of AI agent harnesses, orchestration frameworks, and harness techniques to MCP-compatible agents. The underlying project also publishes a searchable website, an llms.txt file, and JSON data for other forms of access. Its catalog covers projects concerned with agent environments, orchestration, lifecycle management, guardrails, memory, browser use, evaluation, sandboxing, and related areas described in the repository.

The list is intended to support decisions rather than provide a runtime for executing agents. It combines project listings with rankings, capability-oriented filtering, and comparison material. The repository describes the data as refreshed with weekly rescoring, while the displayed project count may change as the list is updated.

How it works

An agent connects to the MCP server and queries the catalog through recommendation functions. The documented examples include recommend and pick_harness; the README indicates additional MCP operations are available, but does not enumerate all of them in the provided material.

Recommendations are meant to account for the model and task being considered. The project specifically presents harness selection as a pairing decision: a harness that performs well with one model may not rank the same way with another. The surrounding guides help turn the catalog into a more detailed choice using factors such as autonomy, recovery behavior, environment, orchestration, lifecycle, and guardrails.

Setup and configuration

The provided material confirms that the repository includes an MCP server, but it does not give an installation command, package name, transport configuration, client configuration example, or required environment variables. Refer to the repository's current agent-facing instructions before configuring a client.

The same catalog is also available through llms.txt and harnesses.json. Those files may be useful when an agent or application needs to consume the data without using MCP, but the excerpt does not specify their schema or update procedure.

Tools and capabilities

The agent-harnesses MCP server supports these documented capabilities:

• Query the curated harness catalog from an agent.
• Request a harness recommendation with recommend.
• Select a harness with pick_harness.
• Use information organized around harness capabilities and decision criteria.
• Consult comparisons covering coding agents, multi-agent orchestration, memory layers, sandboxed execution, browser agents, evaluations, observability, and context files.

The repository also provides a searchable web presentation with one page per harness and filters for capability, autonomy, and recovery. These web features describe the broader reference project; the excerpt does not state that every site function is exposed as an MCP tool.

Limitations and notes

The agent-harnesses MCP server catalogs and recommends projects; it is not described as an execution layer for those projects. It also does not replace testing a candidate harness in the target environment. The repository includes a test-drive guide with a trial process and measurements, indicating that recommendations should be validated against the user's own tasks.

Ranking results should not be treated as universal or permanent. The project notes that harness rankings can vary by model, so changing the model may require revisiting the choice. Counts, scores, and recommendations can also change as the list is refreshed. The provided material does not document authentication, rate limits, persistence, supported MCP clients, or a complete tool schema.