AI agents

Let curxor hand the element you selected straight to an agent. Claude Code, Cursor and Pi are built in; any ACP-compliant agent can be wired up.

Two integration paths, don't mix them up: MCP targets an external editor and hands the selected element's context to the AI there; an Agent targets the curxor side panel and drives the agent directly. They're independent and can run at the same time.

Configuring agents.toml

Agents live in ~/.curxor/agents.toml; the three built-in entries below are written on first start. Edit that file directly afterwards — saves apply on the next send:

toml

[[agents]] id = "claude" name = "Claude Code" command = "npx" args = ["-y", "@agentclientprotocol/claude-agent-acp@latest"] probe = "claude" [[agents]] id = "cursor" name = "Cursor" command = "cursor-agent" args = ["acp"] probe = "cursor-agent" [[agents]] id = "pi" name = "Pi" command = "npx" args = ["-y", "@curxor/pi-acp@latest"] probe = "pi" env = { PI_ACP_KEEP_ALL_SESSIONS = "true" } permissions = false

Field	Meaning
id	Unique key; used to locate the agent when a request is sent
name	Label shown in the “select agent” menu
command / "args"	Command that starts the agent as an ACP server; `args` are its arguments
probe	Binary used to detect whether the agent is installed; falls back to `command`. Claude Code's `command` is `npx` (always present), so `probe = "claude"` is required to detect it correctly
env	Extra environment variables injected into the agent process (optional)
permissions	Whether the agent supports permission prompts. `pi` has no permission dialog, so `false` hides the permission-tier UI (defaults to `true`)

Curxor only probes via probe; it doesn't install agents or manage their sign-in. As long as the binary is on your PATH and the agent is authenticated — via OAuth, API key, enterprise gateway, whatever you use — it just works.

Once installed, the side panel's Settings → Agent settings shows Installed and the agent becomes usable in the select-agent menu. Uninstalled entries are still listed, but model loading fails; see Troubleshooting.

ACP is the key: as long as the command starts as an ACP server over stdin/stdout, Curxor needs no changes — OpenCode, Trae or your own agent all use the same config.

Custom agents

Any ACP-compliant agent can be added with an [[agents]] entry using the fields above. For example, OpenCode:

toml

[[agents]] id = "opencode" name = "OpenCode" command = "opencode" args = ["acp"] # env = { SOME_VAR = "value" } # optional: env vars for the agent process

MCP setup

CLI reference

For AI agents: llms.txt · llms-full.txt