Codex Control Plane MCP

English | Русский

Reliable Codex Desktop automation for long tasks.

codex-control-plane-mcp turns Codex Desktop and codex-app-server into a durable worker that an MCP client can drive safely. Send a task, get an operationId or workflowId right away, poll until the work finishes, approve Plan Mode when needed, then read the final report.

The server handles the awkward parts that thin wrappers usually leave to the caller: app-server startup, thread and turn creation, retry safety, duplicate prompt protection, Plan Mode, approvals, local history, diagnostics, and repair.

OpenClaw and Hermes are first-class clients, but the server is useful for any local orchestrator that needs Codex Desktop to do long-running work without holding one MCP call open for hours.

The short version

MCP client / orchestrator
  -> submit a task or start a Plan Mode workflow
  <- receive operationId or workflowId immediately
  -> poll status
  -> answer approvals or approve the plan
  <- read final report, diagnostics, threadId, and turnId

That gives you a simple contract:

• no multi-hour MCP calls;

• no duplicate Codex turns after a client retry;

• no blind fire-and-forget task submission;

• a local SQLite record of operations, workflows, turns, hooks, and diagnostics.

Related MCP server: deepseek_harness

Why not just call Codex directly?

For a more detailed decision guide, see docs/THIN_WRAPPERS.md.

Current support

• Full live target: Windows with Codex Desktop and codex-app-server.

• Linux and macOS: protocol-only checks for now.

• Local-first: not intended to be exposed as a public network service.

Security model

This is a local-first control plane for trusted Codex Desktop environments.

Do not expose it as a network service without authentication.

Recommended first-run posture:

• use read-only for untrusted repositories;

• use on-request approval when testing new workflows;

• Plan Mode never runs with a read-only sandbox. If a caller requests read-only, MCP raises that turn to workspace-write and reports the adjustment in status output;

• keep state/, logs/, .env, and .codex/ private.

What it does

• Durable async queue for Codex write operations.

• Retry-safe client_request_id handling.

• Active duplicate prompt detection.

• SQLite leases and heartbeats for competing MCP processes.

• Recovery after MCP restart during thread/start or turn/start.

• Durable turn/steer for adding context to an active turn without creating a second turn.

• Durable thread/fork for branching an existing thread, with or without an initial message.

• Plan Mode workflows: start plan, poll, approve, execute, read final report.

• Plan Mode runtime floor: workspace-write, with runtimePolicyAdjusted in status when MCP raises a read-only request.

• Code review workflows through app-server review/start, with polling and final report capture.

• Structured final reports with output_schema.

• Thread lifecycle tools for archive, unarchive, and pollable compaction.

• Workflow goal sync with Codex Desktop thread goals.

• Image and local image inputs for turns that start through turn/start.

• Pending approvals and questions exposed as pollable MCP state.

• Turn interrupts by threadId/turnId, operationId, or workflowId.

• Runtime inventory for models, permission profiles, sandbox readiness, hooks, skills, provider features, account status, usage bands, rate-limit state, and supported app-server methods.

• Health checks, diagnostics, issue analysis, and dry-run repairs.

• MCP-owned hook history in SQLite for search, summaries, and fallback reads.

• Redacted app-server progress journal for deltas, warnings, model reroutes, and token usage.

• Structured MCP errors that automation code can branch on.

Write and control actions go through codex-app-server. The server does not mutate Codex internal SQLite databases or transcript files.

Install

Recommended:

pipx install codex-control-plane-mcp

Or run directly:

uvx codex-control-plane-mcp

From GitHub:

python -m pip install "codex-control-plane-mcp @ git+https://github.com/aresyn/codex-control-plane-mcp.git"

For local development:

git clone https://github.com/aresyn/codex-control-plane-mcp.git
cd codex-control-plane-mcp
py -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install -e ".[dev]"
python -m pytest -q

MCP client config

After installation, generate a config:

codex-control-plane-mcp-admin init --state-db .\state\codex-mcp-state.sqlite3 --projects-root C:\Users\you\Projects

Minimal stdio entry:

{
  "mcpServers": {
    "codex-control-plane": {
      "command": "codex-control-plane-mcp",
      "args": []
    }
  }
}

More copy-paste examples for Claude Desktop, Cursor, VS Code-style MCP clients, local checkouts, installed packages, and central worker mode are available in examples/mcp-client-configs.md.

Run the MCP stdio server:

codex-control-plane-mcp

Or run it as a module:

py -m codex_control_plane_mcp.server

The old openclaw-codex-mcp and openclaw-codex-mcp-hooks commands remain as compatibility aliases for one release line.

Central worker mode

The default inline mode is still the simplest setup: one MCP process can submit and execute operations. For OpenClaw, Hermes, or any setup with several MCP clients, use a central worker instead.

Recommended local shape:

• every MCP client uses the same CODEX_HOME and CODEX_MCP_STATE_DB;

• OpenClaw gateway entries run with CODEX_MCP_EXECUTION_MODE=client;

• one long-running codex-control-plane-mcp-worker process owns codex-app-server, leases, queue slots, and resource locks;

• clients call codex_submit_task, then poll status. They do not execute queued operations themselves.

Worker command:

$env:CODEX_MCP_EXECUTION_MODE = "worker"
codex-control-plane-mcp-worker

Safe observation mode, useful before switching a live gateway:

codex-control-plane-mcp-worker --observe

Concurrency defaults:

CODEX_MCP_MAX_ACTIVE_TURNS_GLOBAL=4
CODEX_MCP_MAX_ACTIVE_TURNS_PER_PROJECT=3
CODEX_MCP_MAX_ACTIVE_TURNS_PER_AGENT=3
CODEX_MCP_MAX_ACTIVE_TURNS_PER_THREAD=1
CODEX_MCP_MAX_ACTIVE_WRITE_TURNS_PER_PROJECT=1
CODEX_MCP_MAX_APP_SERVER_PENDING_REQUESTS=8

For write turns in the same project, pass resource_keys to codex_submit_task. Without them, workspace-write and danger-full-access turns take a broad project write lock. With disjoint keys, the worker may run several write turns in parallel.

New status tools:

• codex_get_worker_status

• codex_get_queue_status

• codex_get_concurrency_status

• codex_get_worker_command_status

codex_get_operation_status also returns queueState, workerState, slotState, and resourceLockState. A running turn has slotState.claimed=true and a slotClaim with the worker id, slot type, and claim time. codex_get_queue_status separates queued work from running turn operations, auxiliary operations, active turn slots, and lock conflicts.

When a workflow is waiting for capacity, codex_get_workflow_status mirrors the nested operation queue state in workflowOperationQueueState. Use nextRecommendedAction="wait_for_worker_slot" for slot pressure and nextRecommendedAction="wait_for_resource_lock" for write lock conflicts. Do not create another operation for the same work while either action is returned.

First setup

The admin helper can generate a fuller client config, install hooks, and run a protocol smoke:

codex-control-plane-mcp-admin init --state-db .\state\codex-mcp-state.sqlite3 --projects-root C:\Users\you\Projects

The command prints a JSON block you can copy into an MCP client config. It does not print secrets or private prompts.

You can also install only the Codex hooks:

codex-control-plane-mcp-hooks install --state-db .\state\codex-mcp-state.sqlite3
codex-control-plane-mcp-hooks status
codex-control-plane-mcp-hooks doctor

The installer backs up ~/.codex/hooks.json, merges its handlers with your existing hooks, stores stateDb as an absolute path, and writes prompts, visible agent progress text, final answers, and turn status into the MCP state DB. Tool calls and command outputs are not recorded by default. Restart Codex after installing or changing hooks.

For turns launched through codex-app-server, the server mirrors the accepted prompt, visible assistant messages, and turn status into the same SQLite history. That keeps search and status reads useful even when app-server does not execute user hooks itself.

Main workflows

Submit a durable task:

codex_submit_task
  -> operationId
codex_get_operation_status(operationId)
  -> queued / running / waiting_for_approval / completed / failed

Use the same client_request_id when a caller retries after a transport timeout. The retry returns the existing operation instead of creating another turn.

Attach screenshots or other image evidence:

codex_submit_task(
  operation_type="start_chat",
  message="Analyze this screen.",
  input_items=[
    {"type": "localImage", "path": ".\\screens\\error.png", "detail": "low"},
    {"type": "image", "url": "https://example.com/screenshot.png", "detail": "high"}
  ]
)

Image inputs are accepted only for operation types that start a new turn: start_chat, send_message, execute_plan, and fork_thread with an initial message. MCP sends the path or URL to codex-app-server, but operation status and diagnostics return only safe metadata such as type, detail, size, extension, and hashes. Binary image content, raw URLs, and full local image paths are not stored in public status payloads.

Steer an active turn:

codex_submit_task(operation_type="steer_turn", thread_id=..., expected_turn_id=..., message=...)
  -> operationId
codex_get_operation_status(operationId)
  -> follows the target turn until completed / failed / interrupted

Use steer_turn only while the target turn is active. For a completed thread, use send_message instead.

Fork a thread:

codex_submit_task(operation_type="fork_thread", source_thread_id=...)
  -> operationId
codex_get_operation_status(operationId)
  -> completed, threadId=<forkedThreadId>

Start work in the fork right away:

codex_submit_task(operation_type="fork_thread", source_thread_id=..., message=...)
  -> operationId
codex_get_operation_status(operationId)
  -> follows the first turn in the forked thread

Use client_request_id for retry-safe fork requests. Without it, each call is treated as a new fork request. threadId in operation status is the forked thread; the source thread is reported in forkState.sourceThreadId.

Manage thread lifecycle:

codex_archive_thread(thread_id)
  -> completed
codex_unarchive_thread(thread_id)
  -> completed
codex_start_thread_compaction(thread_id)
  -> actionId
codex_get_thread_compaction_status(actionId)
  -> running / completed / unknown_after_app_server_exit

Archive and unarchive are audit actions around app-server thread/archive and thread/unarchive. They refuse to run while the thread has an active turn or a pending interaction. Compaction uses its own lightweight actionId because thread/compact/start is asynchronous. Public thread/delete is intentionally not exposed.

Ask for a structured final report:

codex_submit_task(operation_type="start_chat", message=..., output_schema={...})
codex_approve_plan(workflowId, output_schema={...})
  -> operationId / executionOperationId
codex_get_operation_status(operationId)
codex_get_workflow_status(workflowId)
  -> finalReport.text + finalReport.structured

output_schema is passed to app-server turn/start and is tracked by a schema hash in status output. Object schemas must use the strict form required by Codex: set additionalProperties to false. MCP stores the final assistant message as readable text, then parses JSON object output into finalReport.structured when Codex returns valid JSON. Plain text still works and stays available in finalReport.text.

MCP does not extract hidden chain-of-thought and does not store raw tool payloads or command output in final reports.

Drive Plan Mode:

codex_start_plan_workflow
  -> workflowId
codex_get_workflow_status(workflowId)
  -> wait_plan / review_plan / execute_plan
codex_approve_plan(workflowId)
  -> executionOperationId
codex_get_workflow_status(workflowId)
  -> finalReport

Plan Mode has a runtime floor. The public default write policy is still read-only and on-request, but Plan Mode needs a writable workspace on Windows. If the caller or server default resolves to read-only, MCP sends workspace-write to codex-app-server and returns requestedSandbox, effectiveSandbox, and runtimePolicyAdjusted in workflow and operation status.

Mirror a workflow goal into Codex Desktop when the client has one:

codex_start_plan_workflow(goal="Review the migration plan", goal_completion_action="clear")
codex_get_workflow_status(workflowId, refresh_live_goal=true)
  -> threadGoal.syncState + threadGoal.currentGoal

MCP writes a thread goal only when the client passes goal. Managed goals use clear after completion by default. Use set_complete or leave when the goal should remain visible after the workflow ends. Normal workflow polling is passive; use refresh_live_goal=true only when you want MCP to call live app-server goal methods.

Run a Codex code review:

codex_start_review_workflow(thread_id=..., target_type="base_branch", base_branch="main")
  -> workflowId
codex_get_workflow_status(workflowId)
  -> wait_review / read_review_report

Or let MCP create a service thread for a local checkout:

codex_start_review_workflow(cwd=..., target_type="uncommitted_changes")
  -> workflowId
codex_get_workflow_status(workflowId)
  -> reviewThreadId + reviewTurnId + finalReport

Review workflows do not write files by themselves. They run inside the selected Codex sandbox and approval policy. Use client_request_id when a caller may retry the start request after a transport timeout.

Handle approvals and questions:

codex_list_pending_interactions
codex_answer_pending_interaction

Start diagnostics with:

codex_get_runtime_capabilities
codex_health_summary
codex_collect_diagnostics
codex_analyze_issue
codex_repair_issue

Repair actions default to dry_run=true.

Status and diagnostic tools also return agentGuidance and agentGuidanceText when MCP sees a blocker, failed state, stale run, pending interaction, duplicate prompt, auth problem, rate limit, or unsafe recovery loop. Agents should follow agentGuidance.instructions before deciding to retry or stop. If agentGuidance.loopGuard.allowed=false, stop automatic recovery, collect diagnostics, and ask a human. Do not create a new client_request_id after a timeout unless the guidance explicitly says to start a replacement workflow.

For a broken Plan Mode workflow, use retry_workflow_with_runtime_policy. It creates a new workflow with the selected sandbox and approval policy, links it to the old workflow through workflowRetryState, and does not revive the old terminal turn.

codex_health_summary is about current readiness by default. Old stale or orphaned rows are reported in historicalDebt, but they do not make fresh orchestration look broken when the worker, queue, and app-server are currently healthy. Use targeted cleanup for that debt instead of blocking new work.

Status payloads now separate freshness signals:

• operationRowAgeSeconds: age of the durable operation row;

• turnFreshness.lastProgressAgeSeconds: age of the last turn progress event;

• workerFreshness.heartbeatAgeSeconds: age of the worker heartbeat;

• stalenessMeaning="operation_row_age" for the compatibility stalenessSeconds field.

Public status payloads are agent-safe. Operation and workflow status return requestSummary instead of raw request; it contains ids, runtime policy, scheduling intent, input item state, output schema hash, resource keys, and text hashes. It does not include the full prompt, full instructions, raw title, raw image URL/path, exact token counts, raw command output, or private paths. Use your own stored task text plus requestSummary.*.sha256 for correlation.

codex_get_queue_status only recommends wait_for_worker_slot when there is actual queued work blocked by slots. If there are running turns but queueSummary.queued == 0, the queue action is none.

Runtime capabilities

Use codex_get_runtime_capabilities before orchestration or after reconnect. It starts the MCP-owned app-server if needed, calls short best-effort inventory methods, and returns a cached snapshot for five minutes.

In client mode, the client process does not start its own app-server for live inventory. It returns a passive worker-managed snapshot when one exists. With refresh=true, it queues a worker command and returns refreshCommandId; poll codex_get_worker_command_status to read the refreshed inventory.

The response includes:

• model count, default model, hidden flags, input modalities, reasoning efforts, and service tier count;

• permission profiles by id and description;

• Windows sandbox readiness;

• provider capabilities for web search, image generation, and namespace tools;

• hook and skill counts without raw hook commands or absolute skill paths;

• redacted account status, coarse usage bands, and operational rate-limit state;

• supported app-server schema methods with a compact source, version, and hash.

Account inventory is safe to show to an orchestrator. It reports whether Codex is authenticated, the account and plan type, whether an email exists, whether usage data is available, and whether a rate limit or credits issue is visible. It does not return raw email, account identifiers, credit balances, spend limits, exact spend used, daily usage buckets, or exact token counts.

If one inventory method times out or fails, the tool still returns ok=true with runtimeCapabilities.status="partial" and a machine-readable warning in methodResults. Set refresh=true to bypass the cache. codex_health_summary shows a small