Subagent MCP

One agent that plans, implements, and reviews the same work is also grading its own assumptions. Subagent MCP keeps Codex as the main agent and orchestrator, then lets it delegate bounded work to external agent runtimes. Each runtime is an independent model paired with its native harness, so implementation and review can come from a different model, context, and set of assumptions.

Adapters translate each native harness into one normalized lifecycle. The core does not hard-code provider roles or model names. These runtimes supplement Codex's native subagent pool and can use provider quota under an explicit runtime billing policy. Subagent MCP never enables, purchases, auto-reloads, or silently opts into usage credits or paid overage.

Preview: 0.1.0a19 targets Windows. The local MCP, deterministic adapter, package, localhost UI, and Claude Code native-harness integration are ready.

Runtime status

• Claude Code — Ready. It delegates through the native Claude Code harness, keeps model and reasoning choices provider-native, and verifies subscription OAuth identity plus live no-overage evidence before accepting its output. Current provider availability is shown separately in the localhost UI.
• DeepSeek Harness — In development. The current source includes a first native ACP vertical slice. It discovers a standard Windows Node install even when an MCP client filters ProgramFiles, and the source checkout linked by the native ~/.dsh profile without depending on a separate web launcher. Initial provider-backed review proof has passed; broader provider and lifecycle coverage remains in progress. Billing may use credits or unlimited offers the user already authorizes; auto-top-up and overage are never enabled.

Install

Install uv first if you do not already have it:

winget install --id=astral-sh.uv -e

Then install the pinned preview and connect it to Codex:

uv tool install subagent-harness-mcp==0.1.0a19
codex mcp add subagent-mcp -- subagent-harness-mcp serve

Start a new Codex task after registration. You can confirm the installation at any time:

subagent-harness-mcp --version
codex mcp list

If 0.1.0a19 has not reached PyPI yet, install the current checkout instead:

uv tool install .

Open the local UI

subagent-harness-mcp ui

This opens http://127.0.0.1:8765 for settings, health, and read-only activity. It does not require the MCP server to be active. The default foreground command runs until you press Ctrl+C; the page is not an agent chat window.

To keep the UI available after the terminal closes, start the optional managed background process. It remains independent of MCP until you stop it or the Windows session ends:

subagent-harness-mcp ui --background
subagent-harness-mcp ui --status
subagent-harness-mcp ui --stop

Use --background --no-open when you want the service available without opening a browser tab. Subagent MCP does not add itself to Windows login or startup automatically.

Choose another fixed port, or ask the OS for a temporary one, when needed:

subagent-harness-mcp ui --port 9123
subagent-harness-mcp ui --port 0

Background mode requires a fixed port so status and graceful stop target the same loopback service.

On Windows, stop the background UI before upgrading or removing the uv tool so the running Python environment does not hold package files open:

subagent-harness-mcp ui --stop
uv tool install --reinstall subagent-harness-mcp==0.1.0a19

Use it from Codex

After registering the server and configuring a runtime, start a new Codex task and delegate in natural language. For example:

Use Subagent MCP to ask an external agent to review this change, then evaluate its findings independently.

Codex decides what to delegate, observes the result, and keeps the final judgment. Underneath, each adapter maps the same lifecycle to its native harness: spawn, inspect or wait, send follow-up input or interrupt, then close.

Configure DeepSeek Harness (development)

Install and configure DeepSeek Harness normally, then open the Subagent MCP UI and enable DeepSeek Harness. Enter the exact native model as provider-name::model-id; Subagent MCP does not maintain a provider or model allowlist. The adapter uses DeepSeek Harness's native ACP transport, not its web UI.

The primary model is followed by an optional Fallback models (in order) list. Enter one exact model ID per line. Codex moves to the next configured variant only after the current provider explicitly reports exhausted quota or credit (QUOTA_PAUSED); ambiguous failures, timeouts, and crashes are reported without an automatic retry. No model, including Ox Alpha, is selected by default for public users.

Enabling this runtime authorizes the selected route to consume quota from an existing subscription or unlimited offer, or an already funded provider balance. Subagent MCP does not purchase, reload, or increase that balance and cannot verify a promotion or price that the native harness does not expose.

On Windows, the adapter discovers Node from PATH or the standard Program Files installation and follows the native ~/.dsh profile link to the source checkout. Non-standard installations can set SUBAGENT_MCP_DSH_NODE and SUBAGENT_MCP_DSH_SOURCE_ROOT before starting the MCP or UI.

If the MCP controller exits during a DeepSeek turn, restart recovery changes the execution only after read-only process inspection proves that the exact conversation-bound ACP process is gone. It never kills an unverified process.

To keep Codex supervision lean without discarding detail, leave lifecycle responses in their default compact mode and use one agent_wait call with its five-minute default. The MCP waits locally and wakes Codex only for completion, required input, or a timeout. A completed agent keeps its full redacted report in local product state, bounded at 65,536 characters. Compact status returns a short capsule or preview plus its SHA-256 and character count; Codex can use agent_result_read to pull only the hash-bound 4,096-character slices it needs. Transport compression such as gzip can reduce network bytes but does not reduce model tokens after decompression, so Subagent MCP avoids opaque compressed text.

How it fits together

flowchart LR
    C["Codex<br/>Main agent & orchestrator"]
    M["Subagent MCP<br/>Gateway"]
    UI["Localhost UI<br/>Settings & activity"]

    C -->|"stdio MCP<br/>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["More runtimes<br/>via future adapters"]
    end

    M -->|"normalized lifecycle"| R1
    M -->|"normalized lifecycle"| R2
    M -->|"normalized lifecycle"| RN

A runtime may be Claude with Claude Code, a Cursor-supported model with Cursor's harness, Qwen with its native harness, or another adapter. These are examples of the adapter shape, not special cases in the architecture.

Subagent MCP owns the normalized lifecycle, status, redaction, leases, and circuits. Each adapter translates that contract to its native harness without writing shared state directly. See the architecture for details.

What works in this preview

Live provider availability still depends on the user's installed native harness, authentication, selected model, and current provider limits.

Other MCP clients

Point any stdio-compatible MCP client at the installed command:

{
  "command": "subagent-harness-mcp",
  "args": ["serve"]
}

The MCP exposes versioned runtime, project-trust, agent-lifecycle, and workspace tools. Public schemas live in schemas/.

Safety and billing

• Subagent MCP never enables usage credits or changes billing settings.
• Each Claude turn is a bounded live native-harness request. The adapter accepts its output only after claude auth status, the live OAuth init event, and a safe no-overage rate event agree. Missing or unsafe evidence interrupts the request and discards its output.
• Claude exposes that rate evidence only on the live stream, so this guard can consume included subscription quota. It cannot inspect or change Claude's account-level usage-credit toggle; subscription-only users must keep usage credits disabled in Claude. Subagent MCP never turns them on.
• Missing local model, workspace, or session configuration blocks launch. Live identity or rate mismatches interrupt before output is accepted. A configured fallback is selected only after an explicit QUOTA_PAUSED result; ambiguous failures never trigger another paid request.
• Provider Refresh is a no-model preflight. It never launches a canary or task; when a native harness cannot expose pre-turn quota evidence, the UI reports Unknown instead of spending provider quota to manufacture an answer.
• Provider model IDs and reasoning settings remain native, opaque values.
• Product data stays in explicit local config, state, and data roots. Optional client registration uses the client's official command and verifies the exact entry instead of directly rewriting unrelated configuration.
• Native transcripts remain owned by the native harness. Treat agent output as untrusted advice and verify it before applying changes.

Read the full threat model and report vulnerabilities privately as described in SECURITY.md.

Development

CONTRIBUTING.md contains the deterministic test workflow and adapter guidelines. Subagent MCP is released under the MIT License.