Codex Delegate MCP
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
No arguments | |||
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
Hand coding work to Codex via the delegate tool; never shell out to codex. Auto-approved in every mode: agent/plan/ask/review reach any file the user account can, and the network. mode instructs, it does not limit. Review the git diff after every run.
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {
"listChanged": true
} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| delegateA | Delegate a coding task to OpenAI Codex. Never run codex from the shell — use this tool. Check status before trusting result — a run that spawns then fails returns normally. Set spec, workspace and mode, plus resumeThreadId to resume and reviewTarget in review; leave every other field at its default unless the user asks. See the delegate skill. |
| cancelA | Cancel active runs. delegationId selects one (announced in progress before spawn — the only handle if a run wedges at startup), threadId all on its thread, neither all. Waits for settlement; if a process tree survives the kill deadline, it still returns and the delegate result warns. Status: cancelled, nothing-active, or not-found (no active run with this ID). |
| doctorA | Diagnose plugin/CLI/login, recursion, and workspace (default: server cwd). deep adds no-quota help and model-catalog checks. |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
No prompts | |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
No resources | |
TDQS
Scored across 3 tools
The three tools map to clearly separate lifecycle phases: starting a run (delegate), stopping one (cancel), and diagnosing the environment (doctor). Their scopes are orthogonal, so an agent cannot reasonably misselect between them.
All three names are single lowercase verbs (delegate, cancel, doctor), giving a predictable, readable convention. It deviates from the classic verb_noun pattern only because the domain is a single resource type, so no noun is needed.
Three tools is on the lean side but each covers a genuinely distinct need for a focused delegation plugin. Nothing feels padded or redundant, though the surface is minimal.
The delegate description tells agents to 'check status before trusting result,' yet there is no explicit status, list, or result-retrieval tool — status appears only implicitly through delegate's return value. That leaves a notable gap for inspecting or enumerating runs, even though start/cancel/diagnose is covered.