Subagent MCP server
io.github.Thang1710/subagent-mcp·v1.0.31Let Codex orchestrate external coding agents through their native harnesses.
0 stars142 downloads/wk
Reviews
Write oneNobody has reviewed Subagent MCP yet.
If you have run it, two minutes of your experience saves the next person an afternoon.
Subagent MCP tools (15, 3 write)
write = sends, deletes, buys or postsRead from the package source without running it. The installed server may list more.
agent_closeClose a logical conversation and return compact status by default.
agent_interruptInterrupt the current execution and return compact status by default.
agent_result_readRead one hash-bound slice of a terminal result without calling a provider.
agent_sendwrite actionContinue a native session, optionally relaying one hash-bound result.
agent_spawnwrite actionCreate one conversation.
agent_statusRead compact status, or the full envelope when explicitly requested.
agent_waitWait locally up to four minutes. Expiry returns running, does not interrupt, and means continue waiting.
project_scanScan project executable content without changing trust.
project_trustTrust or revoke exact approval-gated project path and hash entries.
runtime_authenticateOpen one user-approved native sign-in flow in the system browser.
runtime_canarywrite actionRun the separately gated live harness canary for one exact adapter pair.
runtime_checkCheck locally by default; refresh asks the harness and never guesses reset time.
runtime_configureApply a user-approved runtime settings patch with revision protection.
runtime_listList runtimes, health, priority, ordered model fallback, and circuits.
workspace_releaseSafely release an owned workspace after the approval gate.
Public scan report
scanner v0.1.10 · 2026-10-03 · same rubric, same numbers if you re-run it
- Code scan28 source files scanned25/25
- –Live reliabilityno gateway calls yet and no remote to proben/a
- –Tool poisoningtools not inspected (local package is not executed); not countedn/a
- Auth qualitylocal package, no credentials required12/15
- Maintenancelast push 10 days ago15/15
- Maintainer identityregistry namespace matches repository owner; GitHub account older than a year8/10
How Subagent MCP compares
Against every graded MCP server, from our own scans and uptime checks. Updated after each scan run.
- Safety score
- 92/100
- higher than 93% of 48,243 servers
- Tools
- 15
- the typical server has 7
- Tools that change things
- 20%
- typical: 0%. Fewer write tools means less an agent can break
- Downloads a week
- 142
- more than 70% of 12,306 servers with a package
What its tools let an agent do: 8 read · 2 change · 5 run or deploy
Release history: 2 versions in the registry since 2026-09-13, latest 2026-09-24, typically one every 11 days.
Install Subagent MCP in Claude Code, Cursor or VS Code
claude mcp add subagent-mcp -- uvx subagent-harness-mcp
What the publisher says
From the Subagent MCP repository's README, as published. We do not edit it. Read it on GitHub
Subagent MCP
<!-- mcp-name: io.github.Thang1710/subagent-mcp -->
Independent harnesses. One Codex orchestrator.
When one model plans a change, implements it, and reviews it, the reviewer shares the author's context and blind spots. It can end up confirming its own plan instead of testing it.
Subagent MCP keeps Codex as the main agent and final decision-maker while delegating bounded work to external agent runtimes. Each runtime is a model paired with its native harness, so Codex can get implementation or review from an independent model with different context and assumptions.
That expands Codex's effective sub-agent pool and can use provider quota you already have. Subagent MCP never enables, purchases, auto-reloads, or silently opts into usage credits or paid overage.
Adapters translate every native harness into the same lifecycle: delegate, observe, steer, and close. The core hard-codes no provider role or model name.
Stable: 1.0.30 targets Windows. The MCP, package, localhost UI, and
Claude Code and DeepSeek native-harness integrations are ready.
Runtime status
model and reasoning settings, subscription OAuth identity, and live no-overage evidence before accepting its output. Project/local context and exact auto-compaction-trigger attestation remain explicit capability gaps.
- Claude Code — Ready. Uses the native Claude Code harness, provider-native
harness-published model catalog for bounded tasks. Resume after an MCP restart, exact provider quota evidence, interactive input, and declared MCP remain explicit capability gaps. A provider-retired model route fails terminally and is never replaced without the user's explicit selection.
- DeepSeek Harness — Ready. Uses its native ACP transport and
native login only and allows no credits, paid overage, or model fallback. Version 1.0.30 advertises and accepts reporead only. Bounded writing remains In development and workspacewrite requests fail before a native process starts. Other explicit gaps are terminal/test/Git execution, network/web/browser, MCP/plugins/hooks, nested agents, native worktrees, restart recovery, macOS/Linux support, and exact pre-request quota.
- Grok Build — Ready for read-only review. The Windows adapter uses cached
Runtime behavior comes from adapters rather than provider-specific branches in the core.
Quick start
1. Install
Install uv if needed, then register the exact isolated release and start its background UI:
winget install --id=astral-sh.uv -e
codex mcp add subagent-mcp -- uvx --isolated --from subagent-harness-mcp==1.0.30 subagent-harness-mcp serve
uvx --isolated --from subagent-harness-mcp==1.0.30 subagent-harness-mcp ui --backgroundStart a new Codex task after registration.
2. Configure runtimes
Open http://127.0.0.1:8765 in a browser. If the background UI was stopped, start the same exact release again:
uvx --isolated --from subagent-harness-mcp==1.0.30 subagent-harness-mcp ui --backgroundThe settings and read-only activity UI stays on the fixed loopback port and does not depend on an active MCP connection.
If a runtime shows Sign in required, click Sign in. Its native harness opens the operating system's default browser; Subagent MCP never asks for or stores credentials. Complete login there, then click Refresh. A status check never opens a browser by itself.
Current and recent activity is listed by external agent. Select a row to inspect its model, native harness, workspace, permissions, write set, current stage, lifecycle, elapsed time, and redacted terminal result. Prompts, transcripts, hidden thinking, and raw provider events are never displayed.
3. Delegate
Ask Codex in plain language:
Use Subagent MCP to ask an external agent to review this change, then
evaluate its findings independently.
Shortened. The full README is on GitHub.
Nothing above is checked by us. What we check is on the safety report.
Subagent MCP: common questions
- Is Subagent MCP server safe?
- Yes, by our scan: it is graded A (92/100). Read the Subagent MCP safety report
- How do I install Subagent MCP?
- It runs on your machine. Copy the Claude Code, Cursor, VS Code or Claude Desktop config from the install section.
- Does Subagent MCP need an API key?
- Not as far as the registry entry and our scan can tell: no credentials are declared or required.
- Is Subagent MCP maintained?
- The last commit was 42 days ago (2026-09-24). The latest release is v1.0.31.
- What can I use instead of Subagent MCP?
- Servers from other publishers that do the same job: Claudexor MCP server.
- How does Subagent MCP compare with other MCP servers?
- safety score 92/100 (higher than 93% of 48,243 servers); tools 15 (the typical server has 7); tools that change things 20% (typical: 0%. Fewer write tools means less an agent can break). Full comparison.
Alternatives to Subagent MCP
Same job from other publishers: the closest match first, then the best rated.
- ClaudexorRoute coding work across Claude Code, Codex, Cursor, and OpenCode with shared context.not reviewedEstablishedA