Coding agents

Let Claude Code, Cursor, Codex, and other coding agents run H agents and write SDK code.

Holo MCP bridge

holo mcp is an MCP server on your machine, part of the holo CLI. The host starts it over stdio and calls it to run any agent the way holo run does, including h/holo-desktop on this machine.

For cloud agents with no local install, connect the coding agent to the MCP server instead.

Install into a host

Sign in first. Hosts start holo mcp in the background, so they cannot open a browser sign-in.

holo login
holo mcp install list      # supported hosts, and which are detected
holo mcp install cursor    # one host
holo mcp install           # every detected host

Common host IDs: claude-code, claude-desktop, codex, cursor, vscode, windsurf. holo mcp install list shows the rest.

The entry it adds uses your saved key. Restart the host. For the US region, install with holo --base-url https://agp.hcompany.ai mcp install <host>.

Codex stops a tool call after 60 seconds. Set tool_timeout_sec = 1800 under [mcp_servers.holo] in ~/.codex/config.toml.

For Cursor and other hosts with a config file, the entry is:

{
  "mcpServers": {
    "holo": {
      "type": "stdio",
      "command": "/absolute/path/to/holo",
      "args": ["mcp"]
    }
  }
}

Settings you add by hand:

• Apps launched from the Dock or Start menu do not inherit shell exports, so set environment variables in the host’s MCP env block.
• For local sessions on your own model server, add HAI_AGENT_RUNTIME_BASE_URL and HAI_AGENT_RUNTIME_MODEL there.
• Claude Code, Codex, and Grok Build drop these settings when you run holo mcp install again. Add them back afterwards.

Tools

run_agent(task: str, agent: str = <your default>, local: bool = false, model: str = None) -> str
list_agents(search: str = None) -> str

run_agent starts a session, reports each step to the host, and returns the answer.

• agent is a name from list_agents, like h/web-surfer-pro for the web or h/holo-desktop for this machine.
• local runs the agent loop on this machine, like holo --local.
• model replaces the agent’s model, like -o agent.model=....

Your ~/.holo files apply to every session.

The tool call blocks until the task completes or hits its budget. The server runs one task at a time, so parallel calls wait.

Write a good task

The agent does not see the host conversation. Each task must say which app, workspace, account, and person, and when it is done:

# Good
Open Slack in the Acme workspace and send a DM to Sarah Chen saying "on my way". Return "sent" after the message appears.

# Too vague
Tell Sarah.

The server gives the host this rule.

Stop a session

Cancelling the tool call in the host cancels the session, but some hosts never send the cancel. For sessions on this machine, use the kill switch or holo local stop. A session the user stopped tells the host not to retry.

Check it worked

Restart the host, then ask:

Use run_agent with the h/holo-desktop agent to open TextEdit and write a short note saying it is connected. Return "done" when the note is visible.

If the host cannot find or start holo:

• Run holo mcp install list, then holo mcp install <host-id> again.
• For Claude Code, run claude mcp get holo.
• Run holo whoami. The key must come from ~/.config/hai/.env, not only from your shell.

More fixes are in Troubleshoot.

Coding skill

The /hai-agents skill writes and runs a TypeScript availability check from one prompt

The hai-agents skill teaches your coding assistant H’s APIs, so it writes working code for sessions, agents, and environments from a plain prompt. One SKILL.md works in Claude Code, Cursor, and Hermes. It loads on its own when you mention H sessions or hk-... keys.

This skill runs in your coding assistant. Agents API skills are instructions you attach to an H agent.

Source: computer-use-agents-demos/skills/hai-agents, also published as a Claude Code plugin in the hai-skills marketplace.

Install the skill

Claude Code

Add the marketplace first, otherwise hai-skills is unknown. In any Claude Code prompt:

/plugin marketplace add hcompai/computer-use-agents-demos
/plugin install hai-agents@hai-skills

Cursor

Cursor 2.4+ reads skills from .cursor/skills/ in your project:

git clone https://github.com/hcompai/computer-use-agents-demos
mkdir -p .cursor/skills
cp -r computer-use-agents-demos/skills/hai-agents .cursor/skills/hai-agents

Then run Developer: Reload Window from the command palette.

Hermes

Hermes reads skills from ~/.hermes/skills/ on startup:

git clone https://github.com/hcompai/computer-use-agents-demos
mkdir -p ~/.hermes/skills
cp -r computer-use-agents-demos/skills/hai-agents ~/.hermes/skills/hai-agents

Then restart Hermes.

Build a use case

Describe the workflow. In Claude Code, use the plugin’s slash command. In Cursor or Hermes, a plain prompt loads the skill. The video above came from:

/hai-agents:hai-agents "Generate TypeScript code that navigates to jacquemus.com, finds the France × Nike football jersey, and checks its availability in sizes S and XXL."

The result is examples/product_availability/src/index.ts. More recipes are in the demos repo.