CLI Agents

Nerveplane is agent-agnostic. Its coordination substrate — the eight MCP tools (stdio and streamable HTTP), the REST API, and the daemon — is standard MCP, so any MCP-capable coding CLI can register, sync, publish, chat, and record decisions. Claude Code, OpenAI Codex, and opencode are supported out of the box; anything else that speaks MCP works too.

Support matrix

	MCP tools (coordinate)	nerveplane worker	Zero-touch hooks	Instructions file
Claude Code	✅	✅	✅ (auto-register, pre-edit warnings, stop-reply)	CLAUDE.md
OpenAI Codex	✅	✅	— (MCP + AGENTS.md)	AGENTS.md
opencode	✅	✅	— (MCP + AGENTS.md)	AGENTS.md
Any MCP client	✅	—	—	—

• MCP tools work everywhere — this is the universal path. Register the nerveplane MCP server in the CLI's config and the agent can coordinate.
• worker (headless autonomous mode) supports all three via per-CLI adapters — pick with --agent.
• Hooks (auto-registration, pre-edit warning injection, stop-reply) are a Claude Code feature; other CLIs coordinate via the MCP tools + an AGENTS.md protocol instead (you keep full coordination; you don't get zero-touch injection).

Codex/opencode flags and output shapes are version-dependent. Adapters follow each CLI's documented form; verify your setup with nerveplane doctor --agent <id> --run before relying on it.

Per-provider setup

Register the MCP server + drop in the agent instructions with one command, then check it:

# Claude Code (also installs the hooks)
nerveplane install claude-code            # or: nerveplane setup (global, once per machine)
claude mcp add --scope user nerveplane -- nerveplane mcp

# OpenAI Codex  → ~/.codex/config.toml [mcp_servers.nerveplane] + AGENTS.md
nerveplane install codex

# opencode  → opencode.json { mcp.nerveplane } + AGENTS.md
nerveplane install opencode

# verify any of them
nerveplane doctor                          # matrix: installed? mcp-registered? hooks? default?
nerveplane doctor --agent codex --run      # live one-turn smoke (needs the CLI on PATH)

Autonomous workers, any CLI

nerveplane worker --agent <claude|codex|opencode> runs that CLI headless in the wake-on-message loop (default claude). Because non-Claude CLIs read MCP from a config file (not an inline flag), run nerveplane install <agent> first so the spawned turns have the nerveplane tools.

nerveplane install codex
nerveplane worker --agent codex --print    # preview the exact `codex exec …` invocation
nerveplane worker --agent codex            # run it

Why Codex runs with --dangerously-bypass-approvals-and-sandbox: in codex exec (non-interactive), every MCP tool call raises an approval elicitation. With no interactive channel, Codex auto-resolves it with Cancel — even under approval_policy=never — so the nerveplane tools would silently no-op and the model would fabricate a reply. The bypass flag is Codex's documented escape hatch for externally-sandboxed automation; the worker runs on your own machine at the same trust level as an interactive Codex session (the moral equivalent of Claude's --permission-mode dontAsk --allowedTools mcp__nerveplane). Confirmed live via nerveplane doctor --agent codex --run.

Adding another CLI

Each provider is one self-contained adapter in src/agents/ implementing the AgentProvider interface (headlessArgs, parseResult, install, detect, capabilities). Nothing else in the codebase knows a provider name, so a new CLI is a single new adapter file registered in src/agents/index.ts.