gpt2agent

MCP server for your ChatGPT account: codex login → ChatGPT Plus/Pro inside any MCP client.

An MCP server that puts your ChatGPT Plus or Pro subscription — every model and the account-tier features below — inside Claude Code, Codex, Cursor, Windsurf, Zed, and any MCP client.

📖 Quickstart · Client setup · Troubleshooting · FAQ · Docs index · Account safety

Status — REST conversations restored (v0.0.15, 2026-09-17)

Checked 2026-09-17 against a live account. Re-check your own account any time with gpt2agent doctor (below).

The Sentinel/Turnstile challenge that blocked conversation tools since 2026-09-08 is now solved via a sentinel bridge (see below). 12 of 16 live-tested tools PASS; 4 have documented limitations.

The sentinel bridge

The upstream Sentinel challenge requires a bytecode-VM Turnstile token that the built-in solver cannot produce. The bridge loads a solver from an owner-supplied directory and mints the full header set in one consistent session. It has retry with backoff (3 attempts) for transient failures.

Setup (one time):

# Place the bridge at ~/.gpt2agent/sentinel-bridge/ and create the marker:
mkdir -p ~/.gpt2agent/sentinel-bridge
touch ~/.gpt2agent/sentinel-bridge/ENABLED

The bridge directory must contain wrapper/reverse/vm.py (bytecode-VM turnstile solver). Dependencies: pip install esprima pillow colorama (bridge internals). It is NOT part of this distribution — see docs/dev/specs/sentinel-vm.md for the from-scratch interpreter spec that will eventually replace the bridge.

Control:

• GPT2AGENT_SENTINEL_BRIDGE=/path/to/bridge — explicit bridge directory
• GPT2AGENT_SENTINEL_BRIDGE_OFF=1 — force the legacy gate path (offline/test)
• The ENABLED marker file beside the bridge enables it persistently

gpt2agent doctor

gpt2agent doctor

Probes each read-only surface and prints a status table plus a one-line summary. It never sends a message, never creates a conversation, and never spends quota.

What it does

gpt2agent exposes 25 MCP tools that forward requests directly to ChatGPT's backend API. No proxy process. No separate account. No platform API key. Your codex login, your token, your quota.

If you already have the codex CLI logged in, setup is zero extra steps — gpt2agent reuses $CODEX_HOME/auth.json (or ~/.codex/auth.json by default) and picks up its background-refreshed token automatically.

Works with Claude Code, Codex CLI, and any client that speaks the MCP protocol over stdio.

Install — one line

curl -fsSL https://raw.githubusercontent.com/robotlearning123/gpt2agent/main/install.sh | bash

That command:

1. Installs the published gpt2agent package via pipx in an isolated environment.
2. Reuses $CODEX_HOME/auth.json (or ~/.codex/auth.json) if you've run codex login — no separate ChatGPT token paste needed.
3. Detects which MCP clients you have (Claude Code, Codex, Cursor, Windsurf, Claude Desktop, Zed) and writes the right config for each, honoring CODEX_HOME for Codex.
4. Drops the Claude Code skills (deep-research + gpt2agent) into ~/.claude/skills/.

Or step-by-step

# 1. Install the package globally (isolated venv)
pipx install gpt2agent

# 2. Register with all detected MCP clients (Claude Code, Codex)
gpt2agent install                          # auto-detect everything

# Want only one client?
gpt2agent install --client claude-code   # or: codex, cursor, windsurf, claude-desktop, zed
# (VS Code & Cline: see docs/clients.md for the manual snippet)

# HTTP transport instead of stdio?
gpt2agent install --transport http --http-port 9000

Or as a Claude Code plugin

/plugin marketplace add robotlearning123/gpt2agent
/plugin install gpt2agent@gpt2agent

This bundles the MCP server registration + both skills in one step. You still need the gpt2agent CLI on PATH (pipx install gpt2agent) — the plugin wires the server (gpt2agent run --stdio) and skills, not the Python package itself.

Browser transport (optional)

pip install "gpt2agent[browser]"

# ~/.gpt2agent/config.toml
[browser]
enabled = true
headed = true                # visible Chrome (recommended; Turnstile needs it)
# profile_dir = "/custom/path"  # default: ~/.gpt2agent/chrome-profile
# timeout_s = 180

First launch opens a visible Chrome window for a one-time login — after that, all 9 conversation tools work through the browser without any human steps.

Manual config (if you'd rather not run install)

Claude Code — add to ~/.claude.json:

{
  "mcpServers": {
    "gpt2agent": {
      "type": "stdio",
      "command": "gpt2agent",
      "args": ["run", "--stdio"]
    }
  }
}

Codex CLI — add to $CODEX_HOME/config.toml (default ~/.codex/config.toml):

[mcp_servers.gpt2agent]
command = "gpt2agent"
args = ["run", "--stdio"]

Setup (manual token paste — only if codex isn't available)

gpt2agent setup

Prompts for a ChatGPT session token (saved to ~/.gpt2agent/token.json, mode 600), detects your plan, and registers gpt2agent with your detected MCP clients over stdio — the same wiring as gpt2agent install. The codex login flow is preferred when available because codex auto-refreshes its token; gpt2agent reloads the selected Codex auth file on mtime change so long calls don't 401 mid-flight.

Tools (25)

Chat & reasoning

Transport modes (every conversation tool)

Image & code execution

Account introspection

Memory & instructions

Codex (cloud agent)

Architecture

Native Python implementation — no proxy. The server calls /backend-api/conversation (SSE) directly using curl_cffi for TLS impersonation. The sentinel bridge mints the full OpenAI Sentinel header set (fingerprint-config p + PoW + bytecode-VM Turnstile token) in one consistent session, with retry-with-backoff for transient failures.

$CODEX_HOME/auth.json (default ~/.codex/auth.json) ← auto-refreshed by Codex
~/.gpt2agent/token.json                            ← manual fallback
        |
   gpt2agent  (stdio MCP server, token reloaded on each call)
        |
   ┌────────────────────────────────────────────────┐
   │ Transport selection (per tool call)             │
   │                                                  │
   │  REST (default)    Browser (browser=True)       │
   │  ↓                    ↓                         │
   │  sentinel_bridge     Playwright + Chrome        │
   │  (fingerprint p +    (real page, one-time       │
   │   PoW + VM token)     login, [browser] extra)   │
   │  ↓                    ↓                         │
   │  curl_cffi → chatgpt.com/backend-api/*          │
   └────────────────────────────────────────────────┘
        |
   25 MCP tools  (chat, agent, DR ×2, GPT chat, image gen,
                  code interpreter, canvas, memory r/w,
                  instructions r/w, codex r/w, account introspect)

Citations

Deep Research replies include inline [N](url) citation anchors when the model performs web searches (knowledge-only answers have no citations; numbering may repeat for multiple URLs). The gpt2agent/citations.py module rewrites raw citeturn… markers (private-use unicode) from the stream's content_references mapping into clickable markdown links, with a Sources section appended.

Configuration

Optional, searched in order: ~/.gpt2agent/config.toml, ./config.toml, ~/.config/gpt2agent/config.toml. Full reference: docs/configuration.md.

[server]
host = "127.0.0.1"   # loopback only; the HTTP transport is UNAUTHENTICATED
port = 9000

[models]
chat     = "gpt-5-6"        # default for chat tool
agent    = "agent-mode"     # default for agent tool
heavy_dr = "gpt-6-pro"      # override slug for deep_research_heavy

[browser]
enabled  = false             # opt-in; requires "gpt2agent[browser]" extra
headed   = true              # visible Chrome (recommended for Turnstile)
# profile_dir = "~/.gpt2agent/chrome-profile"
# timeout_s   = 180

Account safety

See docs/account-safety.md for the full design. Key rules:

• One browser profile, forever — never copy cookies between contexts
• Human pacing — min 20s between turns, burst cap, daily budget
• Challenge = backoff, never brute force
• Temporary chats by default (smaller account surface)

Limitations

• gpt_chat with g-p- prefix GPTs (public/store) returns 422 — the conversation_origin payload was reverse-engineered for g- prefix only.
• chat(astra-wm) returns empty — the model may need special system hints for working-memory mode.
• memory_create_via_chat depends on the model choosing to invoke the memory tool; it doesn't always do so from a plain-text prompt.
• deep_research_heavy depends on the DR connector — check chatgpt.com → Settings → Connectors → Deep Research is enabled.
• Deep Research quota: limits and reset timing are account-reported.
• Account-tier features not yet supported: Sora video, Operator/CUA, voice sessions, Projects, Tasks (write), file upload.
• Requires an active ChatGPT Plus or Pro subscription.

Security & risk — read before you run this

gpt2agent talks to ChatGPT's private backend the way the web app does. That has real consequences; please understand them before pointing it at your account.

• It impersonates the chatgpt.com web client. It uses curl_cffi TLS fingerprint impersonation and a sentinel bridge to pass the OpenAI Sentinel challenge. This is very likely against the OpenAI Terms of Service, and automated/abnormal traffic can get your account rate-limited, challenged, suspended, or banned. Use an account you can afford to lose, keep volume human-scale, and don't rely on it for anything critical.
• The HTTP transport is UNAUTHENTICATED. Use stdio (the default).
• Your token stays local. Read from $CODEX_HOME/auth.json with ~/.gpt2agent/token.json as fallback. Sent only to chatgpt.com.
• PII redaction is limited. Emails, phones, and secret shapes are masked; everything else is returned verbatim.
• GPT2AGENT_RAW_DUMP (debug) writes raw unredacted traffic to the given path. Use mode-600 file names and delete after debugging.

Found a security issue? See SECURITY.md.

Development

python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
pytest                              # 513+ tests, live tests auto-skip (SKIP_LIVE)
python -m ruff check gpt2agent tests # lint

Release

1. Prepare a release PR: bump version in 4 files (pyproject.toml, gpt2agent/__init__.py, .claude-plugin/plugin.json, server.json) + dated CHANGELOG.md entry.
2. Verify: python scripts/verify_release.py
3. After merge, tag the merge SHA and push: git tag -a v$VERSION $SHA -m "gpt2agent $VERSION" && git push origin v$VERSION
4. CI publishes to PyPI (trusted publishing) + creates a GitHub Release.
5. Full runbook: docs/release-validation.md

After the release PR is merged, read its exact merge SHA, prove that commit is on origin/main, check out that reviewed tree, then create and push only the intended annotated tag:

set -euo pipefail
git fetch --no-tags origin main:refs/remotes/origin/main
read -r -p "Merged release PR number: " PR_NUMBER
RELEASE_SHA=$(gh pr view "$PR_NUMBER" --json mergeCommit,state   --jq 'select(.state == "MERGED") | .mergeCommit.oid')
test -n "$RELEASE_SHA"
git merge-base --is-ancestor "$RELEASE_SHA" origin/main
test -z "$(git status --porcelain)"
git switch --detach "$RELEASE_SHA"
trap 'git switch - >/dev/null || true' EXIT
test "$(git rev-parse HEAD)" = "$RELEASE_SHA"
VERSION=$(python -c "import tomllib; print(tomllib.load(open('pyproject.toml','rb'))['project']['version'])")
TAG="v$VERSION"
python scripts/verify_release.py --tag "$TAG"
REMOTE_TAG_SHA="$(git ls-remote --tags origin | awk -v ref="refs/tags/$TAG" '$2 == ref { print $1 }')"
if [ -n "$REMOTE_TAG_SHA" ]; then
  echo "Release tag already exists on origin: $TAG" >&2
  exit 1
fi
git tag -a "$TAG" "$RELEASE_SHA" -m "gpt2agent $VERSION"
git push origin "refs/tags/$TAG"
trap - EXIT
git switch -

If a publish or downstream release job fails, use GitHub Actions' Re-run failed jobs on that same workflow run so it reuses the original build artifact. Do not re-run the whole workflow after any file reaches PyPI.

If a publish or downstream release job fails, use GitHub Actions' Re-run failed jobs on that same workflow run so it reuses the original build artifact. Do not re-run the whole workflow after any file reaches PyPI: Python sdists are not guaranteed byte-reproducible, and the hash guard intentionally rejects different rebuilt bytes for an existing version.

Testing

• Offline: pytest — 513+ unit/contract tests (zero network)
• Live verification matrix: scripts/agent-user-journey.sh <worktree> — 15 cases
• Release emulation gate: scripts/release-emulation-test.sh <worktree> — 11 checks
• Parameter contracts: tests/test_param_matrix.py — 34 cases

License

MIT. See NOTICES for third-party attributions.

Acknowledgments

• lanqian528/chat2api — POW solver (MIT)
• Sentinel bridge technique studied from public reverse-engineering (2026); see docs/dev/specs/sentinel-vm.md for the from-scratch interpreter spec.
• basketikun/chatgpt2api — survey of ChatGPT backend API patterns