SubagentBridge
You can use Simpul/SubagentBridge via MCP stdio to submit coding goals, execute and supervise worker tasks, and inspect results with approval and cancellation controls.
Submit high-level goals with optional model, workspace, commands, and execute flag.
Execute a full supervised PM/Developer/QA lifecycle for a run or a specific task using explicit commands.
Monitor runs and tasks: read summaries, list task status/role/dependencies, and retrieve executor receipts and QA verdicts.
Manage gated actions: list pending approvals and reject actions; approval is not exposed as a model-callable tool.
Cancel active runs and mark their tasks cancelled.
Inspect model token/cost usage, discover local 9Router models/capabilities, and list approved/discovered skills by role or workspace.
Use legacy background-agent helpers to spawn sessions, send prompts, wait for idle, get agent status, list agents, and kill agents.
Limitations: autonomy is off, changes require user approval, workers use the same OS identity by default, and this is not an OS sandbox; do not treat unverified or unexecuted tasks as done.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@SubagentBridge Spawn a background subagent to refactor the auth module and report back."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Simpul
Simpul mengatur tugas coding yang dikerjakan beberapa worker di bawah pengawasan pengguna. Kamu dapat mengirim tujuan, memantau tugas, memeriksa hasil, dan membatalkan pekerjaan dari host yang mendukung Model Context Protocol (MCP).
Nama produknya Simpul. Repository, paket Python, dan pengaturan teknis tetap menggunakan SubagentBridge atau subagentbridge untuk menjaga kompatibilitas.
Rilis publik: v1.1.4. Simpul memerlukan pengawasan pengguna. Worker secara default memakai identitas OS yang sama, sehingga jangan menjalankan kode yang tidak dipercaya seolah-olah terkurung dalam sandbox.
Mulai pakai
Siapkan Python 3.11 sampai 3.13 dan Git. Contoh untuk Windows PowerShell:
git clone https://github.com/rzqllh/SubagentBridge.git
cd SubagentBridge
git checkout v1.1.4
python -m venv .venv
.\.venv\Scripts\python.exe -m pip install -e .Lanjutkan panduan menghubungkan Simpul ke Antigravity IDE. Gunakan path absolut Python virtual environment di konfigurasi host. Server ini berkomunikasi melalui MCP stdio; menjalankannya di terminal tidak membuka aplikasi chat.
Untuk Linux dan konfigurasi lain, baca panduan instalasi.
Related MCP server: Roundtable AI MCP Server
Yang tersedia
Kebutuhan | Dukungan |
Membagi dan memantau pekerjaan | Run, task, worker, antrean, dan status tersimpan lokal |
Meninjau hasil | Bukti hasil, log audit, dan pemeriksaan kesesuaian tugas |
Menjalankan tugas dengan pengawasan | Pemeriksaan izin, batas workspace, dan pembatalan |
Memisahkan pekerjaan Git | Worktree per worker dan pemulihan status |
Menghubungkan host/model | Adapter yang memeriksa kemampuan sebelum eksekusi |
Pengiriman tujuan saja tidak membuktikan tugas telah dijalankan. Tugas tanpa perintah eksekusi yang valid tidak boleh dianggap terverifikasi. Dukungan protokol MCP juga tidak membuktikan semua interaksi pada IDE sudah diuji.
Batas penggunaan
Autonomi tetap dimatikan (
AUTONOMY: OFF). Perubahan berizin memerlukan keputusan pengguna yang sesuai.Model dan worker tidak boleh menyetujui tindakan mereka sendiri.
approve_actiondanapply_code_fixtidak diekspos sebagai tool yang bisa dipanggil model.Isolasi worker berbahaya (E18) dan autentikasi operator independen masih menjadi batas keamanan terbuka.
Kontrol Linux dengan identitas worker khusus bersifat eksperimental. Worktree atau proses terpisah bukan sandbox OS.
Pengujian otomatis tidak menggantikan validasi pada host pengguna atau sertifikasi enterprise.
Baca keamanan, batas kompatibilitas, dan kesiapan penggunaan organisasi sebelum mengizinkan eksekusi.
Dokumentasi
Tujuan | Bacaan |
Instalasi dan tugas pertama | |
Memahami status dukungan | |
Memahami desain sistem | |
Integrasi teknis | |
Operasional dan pemulihan | |
Pengembangan dan rilis |
Indeks dokumentasi memisahkan panduan penggunaan dari referensi teknis. Rilis menggunakan GitHub Releases; riwayat perubahan ada di CHANGELOG.md.
Lisensi dan penghargaan
Simpul menggunakan MIT. Implementasi Python ini berdiri sendiri dan mempelajari pola arsitektur Cindy/Orca; lihat third-party notices. Kesamaan pola tidak berarti seluruh kemampuan atau antarmuka upstream sudah tersedia.
Available Tools
20 toolscancel_runB
Cancel an active run and mark its tasks cancelled.
Inputs: run_id (required), reason.
discover_modelsB
Discover available models and their capabilities from the local 9Router instance.
Inputs: workspace_path. Returns: result
execute_runC
Execute the full supervised PM/Developer/QA lifecycle for a submitted run with explicit commands.
Inputs: run_id (required), commands.
execute_taskC
Execute a specific task in a run through the supervised workflow and return evidence.
Inputs: run_id (required), task_id (required), command.
get_agent_statusD
Legacy helper: session diagnostics.
Inputs: session_id (required).
get_model_usageA
Aggregate token and cost usage across active budget ledgers and sessions.
Inputs: no named parameters.
get_pending_approvalsA
List all pending actions requiring human approval under CHAT_APPROVAL mode.
Inputs: no named parameters. Returns: result
get_runB
Read summary status and task count for a given run.
Inputs: run_id (required).
get_task_evidenceB
Retrieve verified executor receipts and QA verdicts for a specific task.
Inputs: run_id (required), task_id (required).
kill_agentC
Legacy helper: force-terminate a session's subprocess immediately.
Inputs: session_id (required). Returns: result
list_agentsC
Legacy helper: list diagnostics for every known session.
Inputs: no named parameters. Returns: result
list_pending_actionsC
Legacy alias: list pending actions.
Inputs: session_id. Returns: result
list_skillsC
List approved, pinned skills and discovered skills for a specific role or workspace.
Inputs: role, workspace_path. Returns: result
list_tasksA
List all tasks in a run with their status, role, and dependencies.
Inputs: run_id (required). Returns: result
reject_actionC
Legacy alias: reject a pending action.
Inputs: action_id (required), reason. Returns: result
reject_approvalC
Reject a pending gated action.
Inputs: action_id (required), reason.
send_messageC
Legacy helper: queue a prompt for a session.
Inputs: message (required), session_id (required). Returns: result
spawn_agentC
Legacy helper: create a new background subagent session on the given backend runner.
Inputs: session_id (required), workspace_path (required), model, runner, +5 more. Returns: result
submit_goalC
Submit a high-level goal to be decomposed, implemented, and verified.
Inputs: goal (required), model, execute, commands, +1 more.
wait_for_idleC
Legacy helper: block until the session is idle/done/error or times out.
Inputs: session_id (required), timeout_seconds.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 20 tools
The core lifecycle tools (submit_goal, execute_run, execute_task, list_tasks, get_task_evidence, get_run) are reasonably distinct, but the surface contains explicit duplicates: get_pending_approvals vs list_pending_actions and reject_approval vs reject_action are the same operation, and the six 'legacy helper' agent/session tools overlap conceptually with the new run/task tools. The 'Legacy' labels in descriptions help an agent pick the modern path, but they do not remove the redundancy.
Nearly everything follows a snake_case verb_noun pattern (submit_goal, list_tasks, cancel_run, get_model_usage), which is predictable and readable. Minor deviations like wait_for_idle and get_agent_status, plus vocabulary drift between 'agent', 'session', 'run', and 'task' for related concepts, keep it from a perfect score.
20 tools is heavy for the apparent scope, and roughly a third of them (spawn_agent, send_message, wait_for_idle, get_agent_status, list_agents, kill_agent, reject_action, list_pending_actions) are deprecated legacy duplicates or aliases rather than new capability. The set would be better scoped near a dozen tools without the legacy tail.
The goal-to-run-to-task-to-evidence lifecycle, cancellation, approvals listing, model discovery, skills, and usage accounting are covered. However, the approval workflow is one-sided: there is get_pending_approvals and reject_approval but no tool to approve an action, which is a hard dead end under CHAT_APPROVAL mode, and there is no retry/resume for a failed task.
Maintenance
Related MCP Connectors
Command center for developers coordinating AI coding agents across projects.
Coordination hub for AI coding agents: message teammates, ask humans, audit every event.
Multiplayer for the coding agents your team already uses: shared tasks, decisions, skills and tools.
Task manager for AI agents. Orchestrate agents to achieve large goals in unified control plane.
Related MCP Servers
- AlicenseAqualityBmaintenanceEnables Claude Code and Claude Desktop to delegate token-heavy tasks to Antigravity headless subagents, offloading file edits, test runs, and exploration while preserving Claude's context window.1225 npm1MIT
- FlicenseNot gradedqualityAmaintenanceEnables a primary coding assistant to delegate tasks in parallel to specialized local sub-agents, sharing project context and synthesizing their responses into a single result.-
- FlicenseNot gradedqualityAmaintenanceEnables a primary AI assistant to delegate tasks to specialized local AI sub-agents in parallel, share project context, and synthesize their responses from within an IDE.-
- FlicenseNot gradedqualityAmaintenanceEnables a primary AI coding assistant to delegate tasks in parallel to specialized CLI-based AI sub-agents, sharing project context and synthesizing their responses from within an IDE.-