SI

Terminal client for Claude.ai and ChatGPT — replays your own browser session against their internal endpoints and streams replies token-by-token. No API key.

Research tool. It runs under your Claude.ai account: its limits, tiers, and billing still apply. It does not bypass rate limits or paywalls.

Demo

With the proxy running, si/haiku shows up in OpenCode's /models picker like any other provider. Pick it, ask, watch the reply stream in — all through your Claude.ai session, no API key.

Install

git clone https://github.com/two35s/reclaude.git && cd reclaude
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python -m si --opencode   # register the MCP tool with OpenCode (see below)

Login (once)

python -m si --from-chrome   # steal session from your logged-in Chrome (recommended)
python -m si --auto-fetch    # or: launch Chrome, log in, auto-extract over CDP
python -m si --manual        # or: paste org_id + Cookie header by hand

Saves to ~/.config/si/claude_session.json (600).

ChatGPT credentials

python -m si --chatgpt --from-chrome   # steal chatgpt.com session from Chrome
python -m si --chatgpt --manual        # or paste the session-token cookie
python -m si --chatgpt --clear-session # delete stored chatgpt creds

Saves to ~/.config/si/chatgpt_session.json (600). The proxy and MCP server pick these up automatically; a chatgpt/-prefixed model then appears alongside the claude ones. Re-run --chatgpt --from-chrome when requests come back 401.

Chat

python -m si                 # REPL
python -m si --tui           # full-screen chat
python -m si --prompt "..."  # one shot
python -m si --model haiku   # haiku | sonnet (default) | opus | fable …

TUI commands: /new /model [name] /models /help /quit.
REPL commands: same, plus /stealth on|off /cleanup /clear-session.

One server-side conversation per session gives multi-turn memory; stealth mode deletes it on exit so it never appears in your browser sidebar (--no-stealth to keep it).

Proxy

python -m si --proxy                 # OpenAI-compatible API on :8765
python -m si --proxy --port 9000
python -m si --proxy &               # run in the background

The proxy must be running for clients to see the models — if it exits, they disappear until you restart it. (Optional: a KeepAlive: true launchd agent keeps it up across reboots.)

Point Crush, OpenCode, or any OpenAI client at http://localhost:8765/v1.

Streaming (text/event-stream) and non-streaming requests are both supported. Every streamed reply is also written to ~/.config/si/last_response.txt — handy when a client shows you nothing.

On a 429 the proxy returns the error instead of silently downgrading you to a weaker model — retries and model choice belong to the client (OpenCode retries on its own).

ChatGPT models

With ChatGPT credentials stored (see above), --proxy also serves chatgpt/<slug> models — chatgpt/auto, chatgpt/gpt-6, chatgpt/gpt-6-mini, … — through the same /v1 endpoint, alongside the claude ones. One proxy serves both providers at once; /v1/models lists everything.

curl http://localhost:8765/v1/chat/completions \
  -d '{"model": "chatgpt/auto", "messages": [{"role": "user", "content": "hi"}]}'

Multi-turn memory works the same way: send an X-Conversation-Id header (or OpenAI user field) and turns share one ChatGPT conversation. Conversation state is keyed per backend, so reusing one key across si/ and chatgpt/ models keeps them separate. Tool calling, streaming, and the 429-surfaces-as-error behavior all apply. Without stored ChatGPT credentials, a chatgpt/ request returns a clear 400 instead.

Requests carry a fallback account id in the sentinel headers (CHATGPT_ACCOUNT_ID_FALLBACK in config.py) for sessions where the auth endpoint doesn't supply one. It was captured from live testing and is unverified for other accounts — if a fresh free account hits unexplained sentinel rejections, start there.

Tool calling works. OpenAI tools are emulated in text: the proxy appends the tool specs to the prompt, asks the model to emit fenced tool_call blocks, and converts them into real OpenAI tool_calls in the response. Send the results back as tool messages on the next request and the model continues. (Claude.ai's native tool channel keeps the stream open waiting for a result, which deadlocks clients that act only on finished responses — so the proxy doesn't use it.)

Use as a normal OpenCode model

With the proxy running, si/sonnet, si/haiku, and si/opus appear in OpenCode's /models picker like any other provider. Select one and chat normally — tools (read/edit/bash), streaming, and multi-turn all work through your Claude.ai account.

python -m si --opencode prints the ready-to-paste providers block for your global config. Or paste it by hand:

{
  "providers": {
    "si": {
      "name": "si",
      "package": "@opencode/ai/providers/openai-compatible",
      "settings": { "baseURL": "http://localhost:8765/v1" },
      "models": {
        "sonnet": { "name": "Claude Sonnet", "modelID": "si/sonnet" },
        "haiku": { "name": "Claude Haiku", "modelID": "si/haiku" },
        "opus": { "name": "Claude Opus", "modelID": "si/opus" }
      }
    }
  }
}

OpenCode MCP server

Expose si to OpenCode as a native tool (claude_ask, chatgpt_ask) without making it your model provider — OpenCode keeps its own tool-capable model and can call Claude.ai or ChatGPT through si when you ask.

Install it with one command, run from this checkout:

python -m si --opencode

That calls OpenCode's own opencode mcp add, so your existing config (comments included) is preserved. Restart OpenCode, then check opencode mcp list.

Your model can then call claude_ask(prompt, model="sonnet"); model accepts aliases (sonnet, haiku, opus) or a full model id. chatgpt_ask(prompt, model="auto") does the same through ChatGPT (needs ChatGPT credentials stored; model takes a chatgpt.com slug like auto, gpt-6, gpt-6-mini).

{
  "mcp": {
    "servers": {
      "si": {
        "type": "local",
        "command": ["/absolute/path/to/si/.venv/bin/python", "-m", "si.mcp_server"],
        "environment": { "PYTHONPATH": "/absolute/path/to/si" }
      }
    }
  }
}

PYTHONPATH is what makes the module importable from any working directory, so no cwd is needed.

Check the handshake standalone:

printf '{"jsonrpc":"2.0","id":1,"method":"tools/list"}\n' | .venv/bin/python -m si.mcp_server

Notes:

• Shared conversations. All claude_ask calls reuse one Claude.ai conversation, and all chatgpt_ask calls one ChatGPT conversation, for the server's lifetime, so follow-ups keep context. Neither is stealth-deleted and both appear in the respective sidebar until you remove them.
• Diagnostics on stderr. si's progress/error output goes to stderr, so it never corrupts the MCP protocol stream on stdout.
• Limitations. Text-only passthrough — no image output, and no tool calls come back. Requests run under your own accounts, so plan limits apply. One call can take up to the HTTP timeout (120 s), longer under rate limiting. chatgpt_ask errors cleanly when ChatGPT credentials are missing or the turn is refused (quota, rate limit).
• Troubleshooting. Session invalid or expired → re-run python -m si --auto-fetch (Claude) or --chatgpt --from-chrome (ChatGPT); the next call picks up the new credentials (no restart needed). If the tool is missing, check OpenCode's MCP logs and that the venv path in command is absolute and correct.

More flags

--batch FILE one prompt per line → ~/.config/si/batch_results.json · --jailbreak FILE prompt from file · --org-id/--cookie inline creds · --clear-session wipe creds · --login save a Chrome profile without extracting · --opencode register the MCP tool with OpenCode and print the provider config.

If it breaks

Security

Your sessionKey and ChatGPT session token are passwords: never share, upload, or commit the credential files. --clear-session (and --chatgpt --clear-session) when done; if exposed, log out in the browser to invalidate them.

The ChatGPT backend drives chatgpt.com's internal API with your own browser session. That is against OpenAI's terms of service and can get the account limited or banned — use a throwaway account, at your own risk.

Dev

pip install -r requirements-dev.txt
PYTHONPATH=. .venv/bin/python -m pytest -q   # tests
.venv/bin/python -m ruff check .             # lint
.venv/bin/python -m ruff format --check .    # formatting

License

Research and educational use only. No warranty.