SubagentBridge

An MCP server that lets a primary agent inside Antigravity IDE delegate coding tasks to background subagents across multiple model backends — with automatic crash recovery, guaranteed process cleanup, and a real Human-in-the-Loop approval gate for destructive actions.

Architecturally inspired by antigravity-mcp by Inferno-Aditya (MIT License). No code from that project is reused directly — see THIRD_PARTY_NOTICES.md for full attribution.

Why

A review of antigravity-mcp's source found it didn't fully deliver on its own docs: timed-out subagents were left running, failed runs weren't retried, a workspace-boundary check could silently fall through, and the advertised "Human-in-the-Loop" safety gate didn't actually exist in code. SubagentBridge fixes all four and adds first-class support for GPT-5.6 (Sol/Luna) and DeepSeek V4 Flash as additional runners. Full rationale in docs/PRD.md.

Status

🚧 Planning stage — see docs/ for PRD, architecture, weakness tracker, API reference, runner spec, security notes, and test plan. Implementation not yet started.

Planned Tools

Tool	Purpose
spawn_agent	Start a new background subagent session
send_message	Send a follow-up instruction to a session
wait_for_idle	Block until a session finishes or times out (kills the process on timeout)
get_agent_status / list_agents	Inspect running/finished sessions
get_model_usage	Token/cost accounting per session
apply_code_fix	Direct surgical find-replace patch, workspace-bounded
kill_agent	Force-stop a session
approve_action / reject_action	Resolve a pending HITL approval request

See docs/API_REFERENCE.md for full signatures.

Supported Runners

• agy — Antigravity IDE's own agent
• claude — Claude Code
• gpt_sol — GPT-5.6 Sol/Luna (planned)
• deepseek — DeepSeek V4 Flash (planned)

Setup (planned)

pip install -r requirements.txt

Add to Antigravity's mcp_config.json — see mcp_config.example.json.

License

MIT — see LICENSE. Third-party attribution in THIRD_PARTY_NOTICES.md.