chatgpt-pro-mcp

chatgpt-pro-mcp is an experimental guarded MCP bridge for Codex/open-ultrawork ProResearchJobV1 workflows.

It is intentionally not a ChatGPT API wrapper and does not expose ChatGPT Pro as a synchronous /v1/responses model. It exposes a narrow async job interface so Codex can submit bounded research work, poll status, fetch a report artifact, and then run Codex-side source verification / claim promotion gates.

Current state: fake adapter + guarded host-side Web adapter. The Web adapter can classify a logged-in ChatGPT page, submit a bounded ProResearchJobV1 only when explicitly enabled, return early for async work, later reattach/fetch the result, and import it as ProResearchResultV1. It still does not expose ChatGPT as a synchronous /v1/responses model, and it does not claim Deep Research equivalence unless the UI/session actually provides that behavior and Codex verifies the returned sources/claims.

Why this exists

Codex/open-ultrawork needs a high-strength research counselor lane:

Codex/open-ultrawork
  -> ProResearchJobV1
  -> chatgpt-pro-mcp async job
  -> guarded ChatGPT Pro Web/App UI lane
  -> report + sources + claim ledger
  -> Codex source verification / supported-unsupported classification / next tests

The user should not be dragged into mechanical copy/paste unless login, CAPTCHA, terms, billing, or another human-only event blocks automation.

Tools

• get_bridge_status
• get_web_preflight
• submit_deep_research
• get_deep_research_status
• fetch_deep_research_result
• cancel_deep_research

This MCP server intentionally does not expose raw browser tools like goto, click, type, read_dom, or cookie/session export.

Run locally

npm ci
npm test
node scripts/selftest.mjs
npm run doctor
node src/mcp-server.mjs

npm run doctor runs unit/contract/security tests, the fake selftest, redaction scan, and Web preflight. Web preflight is advisory: real submission still requires a manually logged-in dedicated browser profile and explicit submit opt-in.

CLI smoke:

node src/cli.mjs web-preflight
node src/cli.mjs submit '{"schema":"ProResearchJobV1","title":"MCP smoke","question":"Research MCP async bridge patterns with primary sources.","privacy_classification":"public","output_requirements":{"language":"zh-Hant","format":"markdown","citations_required":true,"include_uncertainty":true}}'

Fake UI fixture tests

The sandbox includes local fake ChatGPT HTML fixtures for completed/running/login/terms/billing states. These tests validate the state-machine and fail-closed behavior without opening a real browser or touching a real ChatGPT account.

Docker sandbox

Docker is optional. If you want containerized contract/security tests:

docker build -t chatgpt-pro-mcp:sandbox .
docker run --rm chatgpt-pro-mcp:sandbox npm test

For this project, Docker is suitable for schema/security/fake-UI tests. Real ChatGPT Web/App automation will likely need a host-side browser profile adapter because ChatGPT session state must not be baked into a container image or public repo.

Safety rules

Permanent hard-deny actions:

• login / signup / OAuth authorization
• password / OTP / passkey entry
• upgrade / subscribe / purchase / checkout / payment / billing
• accept terms / privacy policy
• enable connector / install plugin
• change settings
• read/export/delete chat history

If encountered, the bridge returns waiting_user or blocked and fails closed.

V0 intentionally rejects file attachments. Put only public-safe, minimized context in the question; do not send local absolute paths, secrets, raw logs, or private thread URLs.

See SECURITY.md and THREAT_MODEL.md.

Host-side Web adapter guard

The Web adapter is present but disabled by default. It only runs when all of these are true:

CHATGPT_PRO_MCP_ADAPTER=web
CHATGPT_PRO_MCP_ENABLE_REAL_UI=1

Even then, it first opens the ChatGPT entrypoint and classifies auth/terms/billing/ready states. It returns waiting_user unless CHATGPT_PRO_MCP_ALLOW_SUBMIT=1 is also set. With submit enabled, it sends only the fixed research-job prompt, waits for the current assistant response, imports the result, and marks all claims as needs_verification by default.

Config values are fail-closed: invalid wait-time env vars or non-ChatGPT entrypoints abort before browser control.

Manual session setup for future real smoke

Use a dedicated profile. The script opens ChatGPT in a normal browser window; you complete login/CAPTCHA/terms manually, then close the browser. The bridge never receives your password/OTP.

CHATGPT_PRO_MCP_PROFILE_DIR=$HOME/.chatgpt-pro-mcp/profiles/default npm run web:login-native

Then guarded smoke can run with:

CHATGPT_PRO_MCP_ADAPTER=web CHATGPT_PRO_MCP_ENABLE_REAL_UI=1 CHATGPT_PRO_MCP_WEB_LAUNCH_MODE=native-cdp CHATGPT_PRO_MCP_PROFILE_DIR=$HOME/.chatgpt-pro-mcp/profiles/default node src/cli.mjs submit '{...ProResearchJobV1...}'

Without CHATGPT_PRO_MCP_ALLOW_SUBMIT=1, the adapter classifies the UI and stops before prompt submission.

Note: web:login-profile uses Playwright and may be blocked by ChatGPT/Cloudflare. Prefer web:login-native for manual login because it launches a normal macOS browser with the same isolated profile.

Real Web smoke

After manual login succeeds, a bounded real smoke can be run with explicit opt-in:

CHATGPT_PRO_MCP_ADAPTER=web \
CHATGPT_PRO_MCP_ENABLE_REAL_UI=1 \
CHATGPT_PRO_MCP_ALLOW_SUBMIT=1 \
CHATGPT_PRO_MCP_WEB_LAUNCH_MODE=native-cdp \
CHATGPT_PRO_MCP_PROFILE_DIR=$HOME/.chatgpt-pro-mcp/profiles/default \
node scripts/web-real-smoke.mjs

Expected result: status=completed, marker_found=true, and result metadata similar to:

{
  "deep_research_confirmed": false,
  "author_model": "chatgpt-pro-web",
  "executor_host": "chatgpt-pro-mcp-web-adapter",
  "authority_mode": "research_ui_bridge",
  "sync_responses_model": false
}

This proves Web UI collaboration and information return. It does not prove full ChatGPT Deep Research mode unless the report itself shows source-backed Deep Research behavior and Codex verifies the sources.

Optional Codex MCP registration

To expose the bridge as a Codex MCP server on a local machine:

bash scripts/install-codex-mcp.sh --preflight    # read-only checks
bash scripts/install-codex-mcp.sh              # guarded, no submit
bash scripts/install-codex-mcp.sh --allow-submit  # enables bounded Web UI submissions

The installer updates $CODEX_HOME/config.toml, creates a timestamped backup, and uses the current checkout path. It does not copy browser profiles, cookies, audit logs, or .data/.

Fresh clone path:

git clone https://github.com/JNSlayer2/chatgpt-pro-mcp.git
cd chatgpt-pro-mcp
npm ci
npm run doctor
npm run web:login-native  # human login only, if needed
bash scripts/install-codex-mcp.sh --allow-submit

Every submitted job includes a public-safe plan_id and the prompt title is prefixed with [Codex ProResearch] so the ChatGPT sidebar remains searchable and you can inspect Codex/subagent communication directly.

Async open-ultrawork behavior

For open-ultrawork-style long research jobs, prefer async mode:

CHATGPT_PRO_MCP_RETURN_AFTER_SUBMIT=1

In this mode submit_deep_research returns running as soon as the ChatGPT thread is created. fetch_deep_research_result can later reattach to the private local thread URL stored in .data/, wait with wait_seconds, extract the latest assistant response, and import a ProResearchResultV1. The private thread URL is not returned through MCP and .data/ must remain uncommitted.

Deep Research selection is best-effort:

CHATGPT_PRO_MCP_RESEARCH_MODE=auto          # try if a visible Deep Research control exists
CHATGPT_PRO_MCP_RESEARCH_MODE=standard      # do not try
CHATGPT_PRO_MCP_REQUIRE_DEEP_RESEARCH=1     # fail closed if the control is unavailable

The bridge records whether a Deep Research control was found in result.authority.research_mode. If it is not enabled, the output is still useful as a ChatGPT Pro Web research memo, but it must not be described as verified Deep Research.

deep_research_confirmed is a separate promotion gate and defaults to false; a clicked UI control alone is not enough to promote a report to confirmed Deep Research.

Source verification helper

After a completed job, you can check source reachability:

node scripts/verify-source-links.mjs --task-id dr_<uuid> --data-dir .data

This only verifies link reachability. It does not verify that a claim is supported by source content; Codex/open-ultrawork still owns claim-by-claim promotion.

Local runtime cleanup

The bridge keeps private task records and artifacts under .data/ so later fetch calls can reattach to the same ChatGPT thread. Clean old local job records with a dry run first:

npm run cleanup -- --older-than 30d
npm run cleanup -- --older-than 30d --confirm

Cleanup removes old .data/jobs/* records and matching .data/artifacts/* folders only. It does not touch browser profiles, cookies, audit logs, or ChatGPT history.