agent-bridge

A Claude Code plugin that lets Claude delegate real work to other locally-installed CLI coding agents — and optionally itself — through a shared adapter interface, and hold multi-turn conversations with each of them, all from inside a Claude Code session.

Currently supported: Codex CLI, OpenCode, Qwen Code CLI, Kilo Code CLI, Antigravity CLI (agy). Adding another one is a matter of registering one more adapter in server/index.js, not building a new pipeline.

What it does

Bundles a small MCP server (Node, stdio transport) exposing three generic tools:

• list_available_clis — probes every registered adapter and reports which are actually installed on this machine, plus each one's most recent session where supported. Always called first — nothing is assumed to be installed.
• agent_delegate — run one CLI (cli: "codex" | "opencode" | "qwen" | "kilo" | "antigravity") non-interactively on a task and get its final answer back as text plus a session_id. permission (defaulting to each CLI's most restrictive tier) controls whether it can write files or run unattended commands.
• agent_list_sessions — list a CLI's existing sessions (auto-titled from their first message), most recent first, including sessions started outside this plugin. Not every CLI supports this (Antigravity doesn't — no non-interactive history exists for it, only a last-conversation cache).

Two skills teach Claude when and how to use these:

• delegate-to-cli-agent — general trigger-based skill: fires when the user asks to "check with codex," "get a second opinion," "run this through opencode," etc. For a comparison, it calls agent_delegate once per CLI (always read-only) rather than relying on a fixed pairing.
• /agent-bridge:delegate — a guided, menu-driven flow: detect what's installed → pick a worker pool (any subset of the installed CLIs, optionally plus Claude itself) → pick a session per CLI in the pool → pick how to delegate (a plain task to one or all pool members, a lanes-split job spread across the pool with Claude able to take a lane directly, or a compare across the whole pool).

Requirements

• Whichever of the above CLIs you want to use, already installed and authenticated on your machine. This plugin does not install or configure any of them for you — it only shells out to whichever ones you already have working. An unauthenticated CLI still gets detected as "installed" by list_available_clis (it just checks the binary runs), but agent_delegate calls to it will come back with ok: false and the CLI's own auth error in warnings rather than crashing.
• Node.js (the bundled MCP server runs on Node; dependencies are vendored in server/node_modules so no npm install step is required after installing the plugin).

Install

/plugin marketplace add <this-repo>
/plugin install agent-bridge

Or for local development:

claude --plugin-dir /path/to/agent-bridge

Watching a delegated call live

Every agent_delegate result includes a log_file path. Each call's raw stdout/stderr is streamed to that file in real time as the delegated CLI runs (not just returned at the end), so you can open a second terminal and watch it happen:

# PowerShell
Get-Content -Wait <log_file>

# bash
tail -f <log_file>

Logs live under %TEMP%\agent-bridge-logs\ (one file per call, so concurrent lanes each get their own).

Security notes

• agent_delegate defaults to each CLI's most restrictive permission tier (read-only for all five currently). Delegated CLIs run non-interactively as child processes of Claude Code with whatever filesystem/shell access their permission level allows — treat write-enabled or full-access modes the same as you would any other command Claude runs on your machine.
• There is no dedicated write-enabled "compare" mode — comparisons are built by calling agent_delegate per CLI with permission: "read-only" explicitly, at the skill level, so nothing running a comparison can accidentally write files.
• Prompts and delegated CLI output pass through as plain text; this plugin does no additional sandboxing beyond what each CLI provides via its own permission/sandbox flags.
• Not every generic parameter is confirmed supported by every adapter (e.g. model/agent/fork on some of the newer ones) — where a flag isn't confirmed, the adapter returns a warning that it was ignored rather than guessing at CLI syntax that might do something unintended.

Known platform notes

• Developed and tested on Windows. The bundled server uses nano-spawn specifically to resolve Windows .cmd/.bat CLI shims without needing shell: true.
• Both codex exec resume/fork reject --sandbox/-C — a resumed/forked session keeps whatever sandbox/cwd it was originally created with. The adapter drops those flags automatically on continuation calls and surfaces a warning if you passed them anyway.
• Kilo Code CLI has been observed not exiting its process after a provider auth error (unlike Qwen, which exits immediately) — the call still resolves correctly once timeout_ms elapses, with the auth error captured in warnings, just slower than a clean failure.
• Antigravity has no non-interactive session-history command — agent_list_sessions for it always returns unsupported rather than an empty/wrong result.