Use The Harness from apps, editors and agents

Apps, editors and other agents can drive The Harness, and The Harness can hand work to them.

From	How	Package
Any MCP client: apps such as Cursor, ChatGPT and Claude, agents such as Claude Code, Codex, OpenCode and Pi	`harness mcp`, an MCP server	`mcp-server`
Zed and other ACP editors	`harness acp`, an agent of the Agent Client Protocol	`acp`
Claude Code, as a plugin	the MCP server, two skills and two commands	this repository’s `plugins/the-harness`
The Harness’s own agents	Claude Code or Codex as a sub-agent	`external-agents`

Every way in starts sessions in the session host, so they show in the app and in /sessions of the terminal chat like any other.

The MCP server

harness install mcp-server

Then add it to your client:

Claude Code: claude mcp add harness -- harness mcp

Codex, in ~/.codex/config.toml:

[mcp_servers.harness]
command = "harness"
args = ["mcp"]

Cursor, in .cursor/mcp.json:

{"mcpServers": {"harness": {"command": "harness", "args": ["mcp"]}}}

harness mcp speaks over stdio. harness mcp --http serves streamable HTTP at http://127.0.0.1:8790/mcp (--port to change it), on this computer only; for example, claude mcp add --transport http harness http://127.0.0.1:8790/mcp.

Tool	Does
`run_task`	starts a session on a task and waits at most `wait_seconds` (up to 120): the answer, or the session and its progress
`list_sessions`, `start_session`, `send_message`, `stop_session`	the sessions
`get_session`	a session’s status, answer and last progress
`observe_screen`	a session’s screen now, as an image
`read_events`	a session’s timeline, as records
`list_packages`, `install_package`	the packages (official ones, by name)
`list_library`	what the agent learned

A long task never blocks the client: it returns the session, which the client follows with get_session. When a session waits for an approval (a sensitive step, the user’s real mouse and keyboard, or every program under approvals.ask = "programs": see the permissions), the tools say so, and the user approves it in The Harness: in the app, or in the terminal chat (/sessions, then /approve). No tool approves it.

The ACP agent

harness install acp

In Zed’s settings.json:

{"agent_servers": {"The Harness": {"command": "harness", "args": ["acp"]}}}

Each thread is a session. The editor shows the agent’s messages, each program it runs as a tool call, and asks you when the agent waits for an approval. A prompt ends with the agent’s turn; the thread’s next prompt goes to the same conversation, and cancelling ends the turn, not the conversation.

The Claude Code plugin

/plugin marketplace add EvolvingLMMs-Lab/the-harness
/plugin install the-harness@the-harness

It brings the MCP server, the skills act-with-the-harness and follow-harness-sessions, and the commands /harness <task> and /harness-sessions. Install The Harness and the mcp-server package first.

Claude Code and Codex as sub-agents

harness install external-agents

An agent’s program hands a task to another agent’s CLI, which works headless on this computer in a directory of its own:

a1 = delegate("write a script that renames the photos by date", agent="claude-code")
wait_agents([a1])

agent is "claude-code" (claude -p) or "codex" (codex app-server). Their messages and tool uses show on the timeline as the sub-agent’s lane, what they spend counts against the budget as for The Harness’s own sub-agents, and they stop at their budget. The CLI must be installed and signed in; when it is not, the sub-agent fails with how to install it.

Codex workspace file operations go through the host. Replacing or deleting a file and reading a credential path require a separate decision. Claude Code’s native tools are limited to reading non-credential paths in its workspace and pass through PreToolUse. Shell, native writes, and unauthorized tool channels stay disabled. Unattended runs refuse sensitive calls.