Skip to main content
Glama

Codex Supervisor MCP

A local Model Context Protocol (MCP) bridge that lets a compatible host start, monitor, steer, interrupt, resume, and approve Codex app-server work.

It wraps codex app-server; it does not automate a terminal UI or scrape an IDE.

Capabilities

The server exposes these MCP tools:

Tool

Purpose

codex_start

Start a new Codex thread and turn in an allowed repository.

codex_send

Send a new instruction after the active turn is idle.

codex_steer

Append guidance to the active turn.

codex_status

Read thread state, recent events, the latest agent message, diff, and pending requests.

codex_wait

Long-poll until completion, failure, interruption, or an approval request.

codex_interrupt

Interrupt an active turn.

codex_list_threads

List persisted threads inside configured roots.

codex_read_thread

Read a persisted authorized thread.

codex_list_approvals

Inspect pending app-server requests.

codex_resolve_approval

Accept, decline, or cancel command-execution and file-change approvals.

Related MCP server: Codex Bridge MCP

Requirements

  • Node.js 22 or newer.

  • A current Codex CLI available as codex.

  • Codex CLI authentication already configured.

  • One or more explicit local repository roots.

This project has no npm runtime dependencies.

Verify the package

npm test

The test suite uses a protocol-compatible mock app-server. It does not make model requests or require Codex authentication.

Install with the Codex CLI

Use the MCP server name codex-supervisor. The name must match CODEX_SUPERVISOR_MCP_NAME; the bridge uses it to prevent the child app-server from loading this same MCP server recursively.

macOS or Linux

codex mcp add codex-supervisor \
  --env CODEX_ALLOWED_ROOTS="/Users/you/code:/Users/you/work" \
  --env CODEX_SUPERVISOR_MCP_NAME="codex-supervisor" \
  -- node "/absolute/path/to/codex-supervisor-mcp/src/index.mjs"

Use the platform path-list delimiter between roots. On macOS and Linux it is a colon (:).

Windows PowerShell

codex mcp add codex-supervisor `
  --env CODEX_ALLOWED_ROOTS="C:\src;D:\work" `
  --env CODEX_SUPERVISOR_MCP_NAME="codex-supervisor" `
  -- node "C:\absolute\path\to\codex-supervisor-mcp\src\index.mjs"

On Windows the path-list delimiter is a semicolon (;).

Confirm the registration:

codex mcp list

In Codex, type /mcp to inspect the connected server.

Install with config.toml

Copy and adapt examples/config.toml, then place its contents in ~/.codex/config.toml or a trusted project's .codex/config.toml.

Use absolute paths. Keep the server id and CODEX_SUPERVISOR_MCP_NAME identical.

ChatGPT desktop or the Codex IDE extension

  1. Open Settings → MCP servers → Add server.

  2. Set the name to codex-supervisor.

  3. Select STDIO.

  4. Set the command to node.

  5. Add the absolute path to src/index.mjs as the only argument.

  6. Add CODEX_ALLOWED_ROOTS and CODEX_SUPERVISOR_MCP_NAME=codex-supervisor.

  7. Save and restart the host.

  8. Type /mcp to verify the tools.

Local STDIO MCP servers are not loaded by ordinary ChatGPT web chats. Using this bridge from the web requires a separately deployed, authenticated remote MCP service or hosted plugin.

Typical workflow

Ask the MCP host to:

Use codex_start in /absolute/path/to/repository to implement the requested
change. Use workspaceWrite, keep network access disabled, wait for progress,
show me every approval request before resolving it, and report the final diff
and test result.

The host should follow this sequence:

codex_start -> codex_wait
  approval request -> inspect -> codex_resolve_approval -> codex_wait
  active correction -> codex_steer -> codex_wait
  completed -> codex_status
  later follow-up -> codex_send -> codex_wait

Every start/send/steer/interrupt call returns an eventCursor. Pass it as afterSequence to codex_wait or codex_status to avoid replaying older events.

approvalPolicy accepts the current app-server wire values on-request (default) and untrusted. The legacy values onRequest and unlessTrusted are accepted by the bridge and normalized before the app-server request.

The public approval API accepts decline even when a Codex app-server release advertises only cancel for the request. In that case the bridge uses the safe app-server cancellation response and reports both the requested and effective decisions.

Configuration

Variable

Default

Meaning

CODEX_ALLOWED_ROOTS

Required

Repository roots separated by the platform path-list delimiter.

CODEX_BIN

codex

Native Codex executable path. Windows .cmd, .bat, and .ps1 shims are rejected.

CODEX_SUPERVISOR_MCP_NAME

codex-supervisor

MCP config id disabled in the nested app-server to prevent recursion.

CODEX_ALLOW_NETWORK

0

Set to 1 to allow callers to request network access.

CODEX_EVENT_LIMIT

1000

In-memory event count, clamped to 100–10,000.

CODEX_SUPERVISOR_DEBUG

0

Set to 1 to copy Codex app-server stderr to this server's stderr.

CODEX_APP_SERVER_ARGS

Internal safe default

Advanced JSON array replacing every argument passed to codex.

The default app-server arguments are equivalent to:

-c mcp_servers.<CODEX_SUPERVISOR_MCP_NAME>.enabled=false app-server

Overriding CODEX_APP_SERVER_ARGS removes that recursion guard. Include an equivalent disable override yourself.

Security model

  • CODEX_ALLOWED_ROOTS is mandatory.

  • Paths are canonicalized with realpath; symlink escapes are rejected.

  • Codex receives restricted read access to the selected repository and platform defaults.

  • workspaceWrite limits writable roots to the selected repository.

  • dangerFullAccess is not exposed.

  • Network access requires both CODEX_ALLOW_NETWORK=1 and networkAccess: true on a task.

  • The bridge has no generic, unsandboxed shell tool.

  • Command and file-change approvals must be resolved explicitly.

  • Threads outside allowed roots are denied or filtered.

  • Event payloads are size-bounded before storage.

  • Stored thread paths are re-canonicalized at use time; deleted or replaced repository paths fail closed.

  • Same-thread mutations, approval responses, and retried remote calls are serialized or deduplicated rather than executed twice.

  • Transport errors are recursively redacted and size-bounded before they cross STDIO or HTTP boundaries.

  • Relay and remote-server credentials (BIOTELE_* and CODEX_REMOTE_*) are removed from the child Codex environment.

  • Remote result submissions are HMAC-authenticated, base64url encoded, split into bounded chunks, and verified by length and SHA-256 before use. Encoding protects the transport from content filters; it is not encryption.

The child app-server still inherits non-relay process settings and your broader Codex configuration. Audit other environment secrets, apps, skills, hooks, and configured MCP servers before using it with untrusted code. Environment stripping is not an operating-system security boundary: a child running as the same Windows user can deliberately query user-scoped settings. Use a dedicated Windows account if that threat is in scope.

Supported approval requests

This release resolves:

  • item/commandExecution/requestApproval

  • item/fileChange/requestApproval

Other app-server requests remain visible through codex_status and codex_list_approvals, but the bridge refuses to answer them. This prevents a generic response endpoint from silently granting permissions or supplying sensitive user input.

Persistence and monitoring

Codex owns persisted thread history. The bridge keeps streamed event buffers, latest deltas, and pending request state in memory. Restarting the MCP server clears that transient state, but codex_list_threads and codex_read_thread can recover authorized persisted threads.

Development

npm test
node --check src/index.mjs

Project layout:

src/app-server-client.mjs  Codex app-server JSONL client
src/approval-policy.mjs    Approval-policy validation and legacy normalization
src/event-store.mjs        Bounded event, turn, and approval state
src/security.mjs           Repository-root policy
src/supervisor-service.mjs Codex lifecycle orchestration
src/tool-registry.mjs      MCP tool schemas and validation
src/mcp-server.mjs         Dual-era MCP STDIO transport
src/index.mjs              Entrypoint
test/                      Unit and integration tests

License

MIT

Codex App Server compatibility

Version 1.0.3 removes the deprecated readOnly.access and workspaceWrite.readOnlyAccess fields from turn/start. Current Codex App Server releases use permission profiles when a client needs custom restricted read scopes. The supervisor continues to restrict writable roots to the selected repository and validates every task directory against CODEX_ALLOWED_ROOTS.

Hostinger remote relay

Version 1.2.5 provides a Hostinger-compatible relay for ChatGPT remote MCP access:

ChatGPT -> OAuth bearer JWT -> Hostinger /mcp -> namespace-routed queue
  codex_*  -> outbound Windows local-agent -> Codex app-server
  reeves_* -> outbound Reeves Android agent -> accessibility service

The public /mcp endpoint validates RS256 OAuth access tokens from an external identity provider. The Windows and Reeves agents use independent HMAC credentials only for outbound polling, status, lease acquisition, and result submission. The Hostinger relay never starts Codex and never reads local repositories.

The hosted relay keeps all existing codex_* tools and additionally exposes reeves_status, reeves_tap, reeves_swipe, reeves_type, reeves_back, reeves_home, reeves_recents, reeves_sequence, and reeves_screenshot. The local STDIO Codex registry remains Codex-only. Agent claims are filtered by authenticated key ID; client-supplied routing fields are ignored.

reeves_screenshot returns Android pixels as a standard MCP image content block (image/png with base64 data) together with width, height, capture timestamp, agent id, and byte-length metadata. The Android agent uses the relay-advertised chunked result protocol, so no unusable Android-local path is exposed and each signed HTTP request remains within the relay body limit.

reeves_sequence sends 1 through 50 ordered device actions in one routed job. Android executes tap, swipe, type, Back, Home, Recents, wait, and screenshot actions locally, stops on the first error by default, and returns one final MCP image by default. Results include indexed action outcomes and additive, secret-free relay/Android stage timings. The existing 25-second agent request is a wake-on-enqueue long poll, not a pickup delay; Android immediately starts another claim after each successful result submission and reuses one OkHttp connection pool.

This release also negotiates a supported MCP protocol version, issues a bounded OAuth-subject-bound session, and requires that session on follow-up requests. Retried tool calls are bound to the OAuth subject, MCP session, typed JSON-RPC id, and request hash; terminating a session invalidates its cached or pending work. The release also cleans up cancelled relay work and crashed app-server state, revalidates authorized paths, isolates per-thread events, and redacts bounded nested error data at every public transport.

Version 1.2.5 also reconciles codex_status.latestAgentMessage with the authorized persisted transcript. Fully persisted external Codex completions, including synthesized rollout-* turns, now replace stale bridge-observed messages while incomplete or interrupted transcript tails remain excluded.

Deploy the updated relay before updating the Windows agent. The new relay still accepts legacy one-shot results, while the new agent uses the chunked format only after the relay advertises support.

See docs/REMOTE_DEPLOYMENT.md for Hostinger hPanel steps, DNS for mcp.biotele.mx, Auth0 setup, Microsoft Entra ID setup, ChatGPT web connector setup and recovery, environment variables, local-agent installation, and the threat model.

Available Tools

10 tools
codex_interruptInterrupt Codex turnA
DestructiveIdempotent

Request cancellation of an active Codex turn.

Inputs: threadId (required), turnId.

codex_list_approvalsList Codex approvalsA
Read-onlyIdempotent

List pending app-server requests. This release resolves command-execution and file-change approval requests; unsupported request types remain visible.

Inputs: threadId.

codex_list_threadsList Codex threadsA
Read-onlyIdempotent

List persisted Codex threads whose working directories are inside CODEX_ALLOWED_ROOTS. Results outside those roots are filtered.

Inputs: cwd, limit, cursor, searchTerm.

codex_read_threadRead Codex threadA
Read-onlyIdempotent

Read a persisted authorized Codex thread. The bridge denies threads outside CODEX_ALLOWED_ROOTS.

Inputs: threadId (required), includeTurns.

codex_resolve_approvalResolve Codex approvalA
Destructive

Explicitly accept, accept for the session, decline, or cancel a pending Codex command-execution or file-change approval.

Inputs: decision (required), requestKey (required).

codex_sendSend new Codex turnA
Destructive

Resume an authorized Codex thread when necessary and start a new turn. Use only when no turn is active; use codex_steer for an in-flight turn.

Inputs: prompt (required), threadId (required), cwd, model, +4 more.

codex_startStart Codex taskA
Destructive

Start a new Codex app-server thread and immediately begin a turn in an allowed local repository. Return threadId, turnId, and an event cursor for codex_wait or codex_status.

Inputs: cwd (required), prompt (required), model, effort, +3 more.

codex_statusRead Codex statusA
Read-onlyIdempotent

Read a thread snapshot, current turn state, pending approvals, latest agent message, latest diff, and recent streamed events.

Inputs: threadId (required), maxEvents, includeTurns, afterSequence.

codex_steerSteer active Codex turnA
Destructive

Append guidance to the currently active turn without creating another turn. Supply expectedTurnId when available to prevent steering the wrong turn.

Inputs: prompt (required), threadId (required), expectedTurnId.

codex_waitWait for Codex progressA
Read-onlyIdempotent

Long-poll an active turn until it completes, fails, requests approval, is interrupted, or reaches the timeout. Continue with the returned eventCursor.

Inputs: threadId (required), maxEvents, timeoutMs, afterSequence.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.1/5.0

    Scored across 10 tools

    Disambiguation5/5

    Each tool targets a distinct lifecycle moment: start/send create or resume turns, steer mutates an active turn, status/wait/read observe, interrupt cancels, and approvals have dedicated list/resolve tools. Potentially overlapping pairs like send vs steer and status vs read_thread are explicitly disambiguated in their descriptions.

    Naming Consistency5/5

    All tools share the codex_ prefix and use clear action-oriented verbs, with verb_noun forms for list/read/resolve operations. The naming is consistent, predictable, and uniformly snake_case.

    Tool Count5/5

    10 tools is well-scoped for a Codex supervisor, covering turn lifecycle, observation, thread inspection, and approval handling without unnecessary redundancy. Each tool has a clear role in the workflow.

    Completeness4/5

    The surface covers the core workflow: starting, resuming, steering, observing, waiting on, interrupting, and listing threads, plus approval management. Minor gaps exist around deleting threads or resolving unsupported approval types, but agents can complete primary supervision tasks.

    Maintenance

    ActivityInactive
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers