Subagent MCP

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: 1.0.31 targets Windows. The MCP, package, localhost UI, and Claude Code and DeepSeek native-harness integrations are ready.

Runtime status

• Claude Code — Ready. Uses the native Claude Code harness, provider-native 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.
• DeepSeek Harness — Ready. Uses its native ACP transport and 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.
• Grok Build — Ready for read-only review. The Windows adapter uses cached native login only and allows no credits, paid overage, or model fallback. Version 1.0.31 advertises and accepts repo_read only. Bounded writing remains In development and workspace_write 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.

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 -ecodex mcp add subagent-mcp -- uvx --isolated --from subagent-harness-mcp==1.0.31 subagent-harness-mcp serveuvx --isolated --from subagent-harness-mcp==1.0.31 subagent-harness-mcp ui --background

Start 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==1.0.31 subagent-harness-mcp ui --background

The 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.

Codex chooses what to delegate, observes the result, and keeps the final judgment. Lifecycle responses are compact by default; full redacted reports remain in local product state and can be read or relayed later by hash-bound reference.

For an approval tied to an exact file, pass inputs on agent_spawn (inside task) or agent_send:

[{"path":"docs/specs/review.md","expected_sha256":"<lowercase SHA-256>"}]

The MCP hashes each repository-relative file read-only immediately before the native turn. Status returns the verified hash and configured reasoning attestation; a changed file fails before the external agent runs.

Models and fallback order

Each native harness publishes its own model choices. The UI shows friendly names and an ordered priority stack; exact provider IDs remain available for advanced routes.

Claude choices come from the installed standalone CLI through the public Agent SDK. Discovery resolves native aliases to exact model IDs without sending a prompt, and caches the resulting metadata for up to 60 seconds. A CLI identity change invalidates the cache. Missing authentication or unavailable discovery leaves custom exact-ID input available; it does not change saved model choices. New model names do not require a new MCP release, but a model can still require a newer native CLI. The usual compatibility canary remains separate.

The model picker adds and removes models only through explicit actions. Opening and applying it leaves the existing order unchanged unless you edit it.

runtime_list includes config_revision. After user approval, configure a runtime through runtime_configure using that revision and a new request ID:

{  "request_id": "choose-claude-model-1",  "expected_revision": 7,  "patch": {    "runtimes": {      "claude-code": {        "options": {          "variant.0.model": "claude-opus-5-5",          "variant.0.reasoning.effort": "max"        }      }    }  }}

The example changes an existing variant. For a fresh runtime, supply enabled: true and options.model_priority with the chosen IDs plus its required reasoning options. Settings and MCP use the same validation and revision check. A repeated identical request observes its saved receipt; different input needs a new request ID. An interrupted mutation with no receipt requires inspecting current state before another change. This tool cannot change project trust and never starts authentication, a live canary, or a provider task automatically.

When a provider explicitly reports exhausted quota or credit (QUOTA_PAUSED), Subagent MCP moves that exact model to the bottom for future tasks. It does not retry the failed task. Ambiguous failures, crashes, and timeouts do not reorder models or trigger another paid request.

That failed task stays terminal, but the next explicit delegation checks the provider again live; no cached reset clock or wait-until checkpoint substitutes for that check. When the native harness exposes no safe pre-request quota endpoint, status stays unknown rather than inferred exhausted.

Operational recovery is capped at three actions. Local state work and explicit pre-provider failures may be retried. DeepSeek also retries an explicit temporary upstream HTTP 429 at most three total attempts. A generic provider failure is never replayed automatically: a read-only task returns bounded guidance for a new explicit attempt, while a write task remains non-retryable until its write set and possible effects are reconciled. Quota, credit, billing, timeout, and ambiguous lifecycle failures never trigger another provider call. Native ACP failures include bounded redacted RPC/provider details when the harness exposes them; unknown quota remains a provider error, not exhaustion.

DeepSeek routes may use an existing subscription, unlimited offer, or funded balance that the user authorizes. Subagent MCP never purchases, reloads, or increases that balance.

Review permissions and source consent

agent_spawn sends the task and source read by the native harness to the selected external provider. An omitted or null required_capabilities defaults to ["repo_read"]. Explicit lists, including [], are preserved; write permission is never inferred. runtime_list advertises each manifest's required_semantic_permissions; Claude and Grok require repo_read.

Missing mandatory permissions fail locally as REQUEST_INVALID, before a runtime probe, task-input read, execution claim or provider launch. The response includes recovery.action="repair", reason="declare_required_permissions" and max_attempts=1. Follow next_action: correct the payload with a new request_id only when the required provider/source access is already authorized. An explicit restriction must not be silently replaced. This correction does not permit replaying a failed native execution or retrying quota, authentication, access, policy or ambiguous-launch failures.

The controller should reuse existing authorization for the named provider and source scope. If a spec-only review needs additional private source, ask before dispatch, for example: "May the selected provider read this spec and these named source folders for a read-only review?" Use the host's supported approval or user-question tool; explain any host rejection, keep the question pending, and honor the answer. Do not repeat a rejected call without the required approval. task.inputs verifies file identities; it is not a read sandbox or consent receipt. The MCP server cannot display a question after the host rejects a call before it reaches the server, and chat consent does not grant OS/OAuth access.

Long-running work

agent_wait is a bounded local observation, not a model deadline. If it returns running with wait_policy=continue_while_running, the external agent may still be working or thinking. Keep observing the same conversation; elapsed time alone never triggers retry, fallback, interruption, or another provider request.

Concurrent writers

A write task can declare up to 32 repository-relative file or directory roots in write_set. External writers may run concurrently when their canonical absolute sets are disjoint. Equal paths and parent/child paths conflict; task and lane names do not affect locking.

Omitting write_set gives the execution the whole workspace for backwards compatibility. Each adapter also enforces the normalized paths at its native harness boundary. Inspect runtime_list before creating a write request: write_root_mode=path-prefix supports exact file or directory prefixes, while existing-directory requires an existing directory. The current DeepSeek Harness native session advertises existing-directory and one writable root. Subagent MCP never widens an exact-file scope to its parent directory; use that broader directory only when it is explicitly acceptable, otherwise choose a runtime that enforces path-prefix. Multiple valid directory roots are split into disjoint writer calls. These leases coordinate Subagent MCP executions; they are not an operating-system sandbox for unrelated local processes.

On the current Windows public slice, set cwd to the checkout root and use workspace="current". For a DeepSeek write, pass workspace_write and exactly one existing directory relative to cwd, for example write_set=["Assets/_Project/Core/Scripts/GameSettings"]. An absolute path, missing path, or exact file is not a valid existing-directory root.

On Windows, the DeepSeek native sandbox must materialize a restricted-token ACE, so the current user also needs WRITE_DAC on the selected directory. Subagent MCP checks that capability read-only before ACP or provider work. It does not change ownership or ACLs. A failed check stays CAPABILITY_MISSING; it never falls back to danger-full-access.

How it fits together

flowchart LR    C["Codex<br/>Main agent & orchestrator"]    M["Subagent MCP<br/>Gateway"]    UI["Localhost UI<br/>Settings & activity"]
    C -->|"delegate · steer · observe"| M    UI --> M
    subgraph E["External agent runtimes — adapter-driven"]        R1["Model<br/>+<br/>native harness"]        R2["Model<br/>+<br/>native harness"]        RN["Future runtimes<br/>via adapters"]    end
    M -->|"normalized lifecycle"| R1    M -->|"normalized lifecycle"| R2    M -->|"normalized lifecycle"| RN

Subagent MCP owns lifecycle normalization, status, redaction, leases, and circuits. Each adapter translates that contract to its native harness. See Architecture for the full contract.

Update or roll back on Windows

Switch versions without reinstalling an environment that may still be running. The first command uses the source version; the add/start commands use the target version. This example upgrades 1.0.30 to 1.0.31:

uvx --isolated --from subagent-harness-mcp==1.0.30 subagent-harness-mcp ui --stopcodex mcp remove subagent-mcpcodex mcp add subagent-mcp -- uvx --isolated --from subagent-harness-mcp==1.0.31 subagent-harness-mcp serveuvx --isolated --from subagent-harness-mcp==1.0.31 subagent-harness-mcp ui --background

Start a fresh Codex task after changing the entry. Existing tasks keep their old runtime until they end. Use the same sequence with the exact versions reversed to roll back.

An already-running MCP that reports UPDATE_QUARANTINED cannot hot-load the replacement safely. Do not retry that resident; finish with native fallback and use the new exact registration from a fresh task.

For a one-time migration from a direct subagent-harness-mcp serve entry, replace the registration first and let the legacy task end naturally. After replacing it, close every Codex window once before optionally removing the now-unused persistent tool. Leaving it installed is safe because every new command above uses uvx --isolated.

Subagent MCP does not edit Codex configuration, kill Codex/provider processes, or clear uv caches on its own.

Safety and billing

• Subagent MCP never enables usage credits or changes billing settings.
• Claude tasks can consume included subscription quota. Each task validates the bound CLI, subscription authentication, credential precedence, and control connection, then requires safe rate evidence from the same response before accepting its output.
• Provider Refresh sends no model prompt. If the native harness cannot expose exact rate evidence before a response, status remains unknown rather than inventing a quota result or using a reset clock.
• Fallback occurs only after explicit quota exhaustion. Unsafe or ambiguous evidence never triggers another paid request.
• Native transcripts remain owned by the native harness. Product state stays in explicit local roots, and agent output must be treated as untrusted advice.

Read Security and the Threat model before enabling write access.

Project

• Architecture
• Adapter authoring
• Contributing
• Changelog
• MIT License