agent-bridge

agent-bridge

A bridge between coding agents — expose each agent's CLI as a spawnable sub-agent tool. Today it ships an MCP server (the agent-bridge-mcp binary, under cmd/agent-bridge-mcp) with bidirectional Claude ↔ Antigravity delegation plus an OpenAI Codex sub-agent. (Formerly agy-mcp / agy-gemini.)

A tiny MCP server that bridges coding agents, exposing each as a spawnable sub-agent tool. One binary registers three sub-agent tools plus a list_agents discovery tool and a parallel_agents fan-out tool:

• antigravity_agent — shells out to the Antigravity CLI (agy --print <task>), i.e. spawns an Antigravity (Gemini) sub-agent. Intended to be called from a Claude session.
• claude_agent — shells out to the Claude CLI (claude --print <task>), i.e. spawns a Claude sub-agent. Intended to be called from an Antigravity session.
• codex_agent — shells out to the OpenAI Codex CLI (codex exec <task>), i.e. spawns a Codex sub-agent. Callable from any parent session.
• list_agents — read-only discovery: reports which backends are installed (and, on request, authed or able to serve a request), so a caller can pick its sub-agent set before fanning out. Takes no task — it inspects the host, it doesn't spawn an agent.
• parallel_agents — runs several sub-agent spawns concurrently in one call and returns all their results. Needed because MCP host clients (Claude Code, Codex, Antigravity) dispatch individual tool calls serially, so issuing N separate sub-agent calls runs them back-to-back; parallel_agents fans them out as goroutines inside this server, so wall-clock ≈ the slowest job, not the sum.

A parent agent calls a tool with a self-contained task; this server shells out to the corresponding CLI, lets the child agent perform it, and returns the child's full output. In effect each tool is a spawned sub-agent callable from inside another agent's session.

All three tools share one backend adapter, so the run / timeout / truncation / result header / context-cancel / loop-guard behavior is identical; they differ only in which CLI they invoke and which CLI-specific flags they support. Adding a CLI agent is a single entry in the in-code backend registry, not new code.

Tool: antigravity_agent

Spawns an Antigravity agent (Google's agy CLI), which runs Gemini models.

Param	Type	Default	Notes
task	string (required)	—	The complete, self-contained task/prompt for the agent.
add_dirs	string[]	—	Directories to add to the agent's workspace (absolute paths). Repeated as --add-dir.
working_dir	string	server cwd	Directory the agent runs in (sets cmd.Dir).
timeout_seconds	number	300 (max 1800)	Maps to agy --print-timeout.
model	string	CLI default	Optional; --model <model> when non-empty. agy has no family alias and bakes effort into the model name (e.g. Gemini 3.1 Pro (High)) — list current names with agy models. No separate effort param.
tier	string	—	Optional reviewer tier deep/fast; the server resolves it to a model by matching agy models output at runtime (deep → a *Pro* (High) entry, fast → *Flash* (Medium)), cached. An explicit model overrides it.
mode	string	reason	Access tier: reason (no permission-bypass flag; but agy has no tool-disable flag and does not gate writes, so a reason agent with a writable working_dir can still read and edit files unattended — use a throwaway working_dir to keep it off your files; omitting working_dir is not a guard (agy then runs in the bridge server's own cwd, often your project tree), and --sandbox does not confine writes) · act (edit files in working_dir + run commands via --dangerously-skip-permissions, unattended). No read tier for antigravity_agent.
sandbox	bool	false	Enable agy's sandbox terminal restrictions (--sandbox). Note: despite the name this does not confine the agent's file edits — a write under --sandbox still lands in working_dir (verified), so it is not a "don't touch my files" guard; use a throwaway working_dir for that. Antigravity-only — claude_agent has no sandbox param.

Tool: claude_agent

Spawns a Claude agent via the claude CLI. This is the reverse direction: intended to be called from an Antigravity session so it can delegate to Claude. It mirrors antigravity_agent's semantics. Note: every run shells out to the claude CLI and therefore consumes Claude credits — even reason-only runs.