wca-mcp
Allows coding agents to send project updates and questions to a signed-in ChatGPT account through browser automation, retrieve replies and conversation URLs, and optionally review PDFs, without using the OpenAI API.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@wca-mcp send ChatGPT a progress report and ask what to do next"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
ask-my-chatgpt
Let Codex and Claude Code talk to your own signed-in ChatGPT through a Chrome extension: send project updates, get a second opinion, and keep one conversation you can come back to.
Unofficial — not affiliated with or endorsed by OpenAI or Anthropic.
This project automates the ChatGPT website in your own browser. That may conflict
with OpenAI's terms of use, and it may trigger rate limits, safeguards or
restrictions on your account. Use it at your own risk. Never send passwords,
API keys, tokens, private keys, .env files or personal data — nothing is
filtered before it reaches ChatGPT.
AI-generated project. The code was generated by OpenAI Codex and Claude Code under human direction. It is a practical tool, not a teaching example. Read the permissions section before installing, and test the parts you rely on. Contributions are welcome.
Status: Windows-first developer preview · MIT license · Install · Workflows · Design · Tests · Contribute
Part 1 — What it does
1. A ChatGPT bridge for coding agents (main feature)
Your coding agent works inside the repository. ChatGPT is often where you think about the bigger picture. ask-my-chatgpt connects the two:
Report progress. Codex or Claude Code sends ChatGPT what was built, which tests passed, what is blocked and what comes next.
Get another view. Ask ChatGPT to challenge a design or suggest an experiment; the coding agent evaluates the suggestion before using it.
Catch up later. Open the same ChatGPT conversation on any device and ask "What changed this week, and what is still open?"
Review PDFs. Send an approved PDF and keep discussing it in the same chat.
It works through the ChatGPT website in your normal, signed-in Chrome — no OpenAI API key. Your account's normal limits apply.
The Chrome Companion is the foundation. This unpacked extension is what
types into the ChatGPT page, confirms the message was sent once and reads the
reply back. It talks only to the local servers on your PC (127.0.0.1), using a
random per-machine pairing token. The extra 1point3acres assistant (below) also
uses it; its files are listed separately in
chrome_extension/README.md.
ChatGPT only knows what you send it. Its summaries can be wrong; the repository and its test results stay the source of truth. Regular reports need a scheduler in your coding client — this project never sends anything on its own.
This repository also ships two independent tools. Neither is needed for the ChatGPT bridge, and both stay off unless you start them.
2. Optional: a local web-content collector
Save articles or an in-scope online course as clean Markdown with images. A
local model in LM Studio picks which links matter and
cleans the text; nothing is sent to a cloud model. Jobs are checkpointed, so an
interrupted download resumes without repeating finished pages. Output goes to
Desktop/WebContent by default. It respects robots.txt and never bypasses
login walls, paywalls or DRM.
3. Extra feature: a supervised 1point3acres daily assistant
Extra, not part of the ChatGPT bridge. You can ignore it entirely: it never starts unless you run it, and the bridge doesn't use it. Its code, scripts, extension files and tests are listed in
src/web_content_agent/onepoint3acres/README.md.
A small Windows desktop app for the daily check-in quiz on 1point3acres, using your own signed-in account. A local model proposes an answer with evidence, a supervisor approves it, the extension submits once, and the website's response is recorded.
Related MCP server: TaskQuay
Quick start (ChatGPT bridge)
You need: Windows, Git, uv, Google Chrome signed in to ChatGPT, and Codex or Claude Code. uv installs Python 3.12.
git clone https://github.com/felixchenfy/ask-my-chatgpt.git
cd ask-my-chatgpt
uv sync --python 3.12 --frozen
uv run python -m web_content_agent.chatgpt_pairingIn Chrome, open
chrome://extensions, turn on Developer mode, click Load unpacked and choose this checkout'schrome_extensionfolder.Register the MCP server with your coding agent (use your real path):
Claude Code
claude mcp add --transport stdio --scope user webcontentagent -- uv --directory "C:/path/to/ask-my-chatgpt" run wca-mcpCodex — add to
~/.codex/config.toml:[mcp_servers.webcontentagent] command = "uv" args = ["--directory", "C:/path/to/ask-my-chatgpt", "run", "wca-mcp"] startup_timeout_sec = 30 tool_timeout_sec = 1200In a new agent session, try:
Use ask_chatgpt to ask "Reply with BRIDGE_OK." with thinking_mode "current". Show me the reply and the conversation URL.
Then tell your agent, for example: "Report our project progress to ChatGPT, ask for alternative approaches, and save the conversation URL privately for next time." More recipes, including periodic reports, are in workflows. Timeouts, PDFs and troubleshooting are in the installation guide.
Check it end to end (sends two short synthetic messages to your ChatGPT):
uv run python scripts/chatgpt-live-check.py --confirm-sendIntegrations
Two optional clients reuse the same bridge and Chrome Companion; both send real messages:
Codex plugin "ChatGPT PDF Review": a Codex skill plus a small Node MCP server (
ask_chatgpt,read_chatgpt_conversation) with PDF-review guidance.Claude Code skill
chatgpt-bridge: aSKILL.mdandgpt.pyCLI; link the folder into~/.claude/skills/.
Both find this checkout relative to their own folder; WCA_PROJECT_ROOT overrides it.
Safety and privacy in one minute
The bridge runs on your PC, but the questions and PDFs you choose are sent to ChatGPT. PDFs are only accepted from folders you allow.
Nothing is scanned or redacted before sending. Keep secrets (passwords, API keys, tokens, private keys,
.envfiles) and personal or customer data out of what your agent sends, and tell your agent the same.Account risk: automated use of the ChatGPT website may conflict with OpenAI's terms and can lead to rate limits or account restrictions. Use only an account and a volume of messages you are comfortable with.
The Chrome extension has strong permissions (tabs, scripting, debugger) so it can type into ChatGPT and read the reply. Read privacy first.
The extension and the local servers pair with a random per-machine token. The local servers listen only on
127.0.0.1and refuse requests from web pages.A timeout does not prove nothing was sent. The bridge records every request and blocks blind re-sends; recovery only reads the reply to the original message.
"Reply extracted correctly" does not mean "reply is correct".
Logins, CAPTCHAs and paywalls are never bypassed.
This is an unofficial, independent project. It is not affiliated with, endorsed by or supported by OpenAI or Anthropic. ChatGPT is a trademark of OpenAI, and Claude is a trademark of Anthropic.
Part 2 — For developers
Architecture
flowchart LR
S["Optional external scheduler"] -. "starts authorized reporting" .-> C["Codex / Claude Code"]
C <-->|"MCP over stdio"| P["Python MCP server"]
subgraph PY["One local Python process"]
P -->|"Python call"| B["Local ChatGPT bridge"]
end
B <--> R["Private task records"]
X["Chrome Companion"] -->|"HTTP poll and report (paired, 127.0.0.1)"| B
B -->|"HTTP response"| X
X <-->|"Browser interaction"| G["Signed-in ChatGPT"]The coding agent decides what to send. Python validates the request and records it durably. The Chrome Companion polls the paired loopback endpoint, types and submits once, then reports the reply that belongs to that exact message. A full request sequence, the collector pipeline and the recovery rules are in docs/design.md.
Repository map
Path | What it is |
| MCP server ( |
| ChatGPT task state, duplicate protection, loopback HTTP API |
| Per-machine pairing tokens for the extension |
| Chrome Companion (MV3 service worker); file map separates bridge, shared and extra files |
| Collector pipeline |
|
|
| Extra: 1point3acres desktop app ( |
| Launchers and diagnostics, grouped by feature in scripts/README.md |
| Codex plugin: skill, Node MCP adapter, its own |
| Claude Code skill ( |
| Python unit/integration tests and Node tests for the extension |
Public interfaces
MCP tools (uv run wca-mcp, stdio)
Tool | Purpose |
| Send one question (optional PDF, optional |
| Read-only recovery of an existing reply; never types or submits |
| List likely content links on one public page (no download) |
| Write a |
| Start the background collector; returns a |
| Concise progress; rate-limited to discourage polling |
| Resume an interrupted run / check outputs and hashes |
| Inventory / redacted failure summary |
| Compact snapshot for continuing in a new chat |
CLI (uv run wca --help): doctor, run, login, resume, status,
validate, handoff, new-source.
Loopback HTTP (used only by the extension; token + extension origin required):
ChatGPT bridge on 127.0.0.1:18732 (/command, /before-submit, /observation,
/event, /submitted, /complete); extra 1point3acres app on 127.0.0.1:18731 (/command,
/status, /run, /run-daily, /simulate, /close, /solve, /complete,
/review/*).
Tests
All offline tests run without a ChatGPT account, LM Studio or network access:
$env:PYTHONUTF8 = '1'
$env:PYTEST_ADDOPTS = '-m "not live and not lmstudio"'
.\scripts\test.ps1 # ruff format + ruff lint + pytest + node --test (+ plugin npm test/lint)The Codex plugin suite runs only when integrations/codex-plugin/node_modules exists
(npm ci there first); CI always installs it.
Area | What is covered | Where |
API contracts | Every MCP tool forwards its arguments; CLI commands; HTTP auth, CORS and unknown routes for both local servers; live-check script refuses to send without |
|
ChatGPT bridge | Send-once state machine, durable idempotency across restarts, zero-submission retry rules, throttling, mode correction, PDF policy, port ownership, read-only recovery |
|
Chrome Companion | Composer readiness, submission confirmation, reply extraction, tab handling, thinking mode, render recovery |
|
Collector | Link selection, extraction fallbacks, images, checkpoints, resume, context budget, model lifecycle, full static/course/resume pipelines |
|
1point3acres assistant (extra) | Quiz parsing, solver output, model selection, supervision and one-use permits, records, recovery, desktop-app wiring, extension quiz flow |
|
Integrations | Codex plugin MCP tools, launcher/agent path resolution, cancellation, cache repair, Companion DOM contracts; Claude skill path resolution |
|
Live (opt-in) | Real LM Studio / website / ChatGPT checks |
|
Current totals and the per-file list are in docs/testing.md.
CI runs the same offline checks on windows-latest for every push and pull request.
What is verified, and what is not
Evidence | |
Offline suite | Windows, 2026-10-15: 484 Python passed, 0 skipped (6 live/model tests deselected); 182/182 Companion JavaScript and 260/260 Codex plugin tests passed, plugin ESLint clean. Run by Windows CI on every push |
Live ChatGPT round trip over MCP stdio |
|
Codex as the MCP client | Long-running personal use |
Claude Code as the MCP client | Standard MCP setup documented; a recorded Claude Code session is still welcome |
macOS / Linux | Offline tests pass on Linux; the browser workflow is not validated there |
Contributing
Bug reports, setup reports, synthetic browser fixtures, clearer docs and focused pull requests are all welcome. Read CONTRIBUTING.md for the workflow, and SECURITY.md for private vulnerability reports. AI-assisted contributions are welcome; you stay responsible for reviewing and testing them. Please follow the code of conduct.
Further reading
Background articles: bridge design, local extraction, supervised browser tasks
Maintained by Feiyu Chen. Licensed under MIT. Model weights, downloaded content and third-party services keep their own terms.
Available Tools
11 toolsask_chatgptA
Send one question and optional PDF through signed-in Chrome and return ChatGPT's answer.
The question and result are persisted in an ignored local task file. Supply a canonical
https://chatgpt.com/c/... URL to continue an existing conversation; omit it for a new chat.
PDF paths must be absolute and inside the existing ChatGPT bridge's allowed roots.
Inputs: question (required), thinking_mode, attachment_path, timeout_seconds, +2 more.
create_collectionB
Create a safe WebContent collection and source.yaml without starting a job.
After this, call start_collection only when the user asks to begin.
Inputs: url (required), name (required), strategy, max_items, +1 more.
create_context_handoffA
Compress durable run state into a small snapshot for continuing in a new Chat.
Use once when a Chat becomes long or before a long background wait. It intentionally omits page bodies, logs, screenshots, repeated tool results, and hidden reasoning.
Inputs: run_id (required).
discover_siteA
Inspect one public page and list likely content links.
This never downloads a collection. Use it first for an unfamiliar site, then ask the user which collection they want.
Inputs: url (required), instructions.
get_failureA
Return a redacted failure summary and bundle path for a failed run.
Never returns cookies, authorization headers, credentials, or full logs.
Inputs: run_id (required).
get_run_statusA
Get concise progress for a run without returning full logs.
Use this when the user asks how a download is going. Never call it repeatedly in one response.
Inputs: run_id (required).
list_collectionsB
List collections with source URL, latest status, and content counts.
Use for questions about existing collections or the latest run.
Inputs: no named parameters.
read_chatgpt_conversationA
Read the latest assistant response from an existing ChatGPT conversation.
This is recovery-only and never types, submits, or uploads anything. Supply the canonical
https://chatgpt.com/c/... URL returned by a previous ask_chatgpt call.
Inputs: conversation_url (required), timeout_seconds, expected_message_id, minimum_response_characters.
resume_runB
Resume a failed or interrupted run without repeating completed work.
Use only after get_run_status or get_failure says the run is resumable.
Inputs: run_id (required).
start_collectionB
Start the existing worker as an independent background process.
Returns a run_id immediately. The path must be a source.yaml inside WebContent.
Inputs: source_config (required).
validate_runB
Validate final outputs, hashes, and local image paths.
Use after status reports success.
Inputs: run_id (required).
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 11 tools
The collection lifecycle tools (create/start/get_run_status/resume/validate/list/get_failure) each target a distinct stage, and the ChatGPT pair (ask vs read) is clearly separated. Minor adjacency exists between get_run_status and get_failure, and between start_collection and resume_run, but descriptions disambiguate them adequately.
Every tool follows a consistent snake_case verb_noun (or verb_resource) pattern: create_collection, start_collection, get_run_status, list_collections, resume_run, validate_run, discover_site, ask_chatgpt. No mixed conventions or vague bare verbs.
11 tools is well within the healthy range and each tool maps to a real step in the web-collection workflow or the ChatGPT bridge. No tool appears redundant or padded.
The surface covers discovery, creation, execution, monitoring, resumption, validation, and failure inspection, which is strong lifecycle coverage. Gaps remain: no cancel/stop for a running collection and no delete/update for collections, though agents can work around these.
Maintenance
Related MCP Connectors
Let AI agents work in your own signed-in Chrome, with approvals and verified results.
Let AI agents use your real, signed-in Chrome: tabs, clicks, forms, screenshots and video.
Let AI agents use any website you're logged in to. Sessions stay encrypted on your machine.
Hand a brief from ChatGPT, Claude or Gemini to your own coding agent or Grok Bot. Send only.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceEnables local coding agents to send prompts to a user-authorized ChatGPT webpage through a browser extension and retrieve the final answer via MCP.1Apache 2.0
- AlicenseNot gradedqualityBmaintenanceEnables ChatGPT conversations to directly inspect local workspaces and delegate coding tasks to a local Codex agent, with controlled concurrency, reusable sessions, and task/usage tracking.597 npmMIT
- FlicenseNot gradedqualityCmaintenanceEnables ordinary ChatGPT web Chat to work directly on a local machine over a private MCP tunnel — reading and editing project files, running shell commands and tests, loading local skills, manipulating images, and driving Chrome — without spinning up a separate coding agent or additional model-generation API calls. It also supports multi-step task plans with persisted execution receipts and acceptance checks so results from real runs are returned to the conversation.-
- AlicenseNot gradedqualityAmaintenanceLets AI coding tools drive the real, already-logged-in Chrome browser over local stdio—snapshotting and reading pages, clicking, typing, managing tabs, capturing screenshots, and inspecting network, console, storage, and cookies—without opening a remote-debugging port. It installs as a Chrome extension plus native-messaging bridge, so agents work in the user's authenticated sessions while passing bot detection.1 npmISC