claude-web2api

A small, (almost) zero-dependency proxy that exposes your claude.ai web session as an OpenAI-compatible API (/v1/chat/completions), so OpenAI-compatible clients — in particular opencode — can use your logged-in Claude account (Sonnet, etc.) as a model, including agentic tool use.

The only dependency is curl_cffi, used to impersonate a browser's TLS so Cloudflare lets the requests through. Everything else is the Python standard library.

Disclaimer — This is an unofficial tool that drives the consumer claude.ai web app with your own logged-in session. It is not the Anthropic API and is not affiliated with Anthropic. Automating the consumer app may violate Anthropic's Terms of Service. Use your own account, at your own risk.

What it does

• Serves POST /v1/chat/completions (streaming + non-streaming) and GET /v1/models.
• No key to paste: reads your claude.ai cookies straight from your Firefox profile (sessionKey, cf_clearance, …) and refreshes them as they rotate.
• Agentic tool use: claude.ai web has no function-calling, so tools are emulated at the prompt level — tool schemas are injected and the reply's <tool_calls> block is parsed back into OpenAI tool_calls. Verified with opencode's build agent (read/write/edit/bash).
• Conversation reuse: keeps one claude.ai thread per session and sends only each new turn instead of replaying the whole history every tool step — much easier on your usage limits.
• Native web search + configurable thinking effort (low…max).
• Image attachments: images in a message are uploaded to claude.ai and attached, so Claude's vision works (sent as files, not inlined text).
• Usage meter: GET /usage reports how much of your rolling 5h / 7d limits you've used (claude.ai doesn't expose token counts; this is its own metric).

Requirements

• Python 3.9+
• curl_cffi — pip install curl_cffi (or pip install -r requirements.txt)
• A Firefox profile logged into claude.ai (or paste a session key manually).

Run

COOKIE_FROM=firefox python3 claude_web2api.py
# -> serving http://127.0.0.1:8080/v1

Offline self-test (no network, no account needed):

python3 claude_web2api.py --selftest

Use it from opencode

Add a provider to ~/.config/opencode/opencode.json:

{
  "provider": {
    "claude-web": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Claude (web session)",
      "options": { "baseURL": "http://127.0.0.1:8080/v1", "apiKey": "local" },
      "models": {
        "sonnet": {
          "name": "Claude Sonnet (web)",
          "tool_call": true,
          "limit": { "context": 200000, "output": 64000 }
        }
      }
    }
  }
}

Then select claude-web/sonnet via /models. "sonnet" resolves to the newest Sonnet your account offers.

Optional: start/stop the proxy together with opencode

Wrap the opencode command in your shell rc so the proxy starts on launch and stops on exit:

opencode() {
  local pid=""
  if ! curl -sf -o /dev/null http://127.0.0.1:8080/health; then
    COOKIE_FROM=firefox python3 ~/claude-web2api/claude_web2api.py >/tmp/claude-web2api.log 2>&1 &
    pid=$!
    for _ in {1..25}; do curl -sf -o /dev/null http://127.0.0.1:8080/health && break; sleep 0.2; done
  fi
  command opencode "$@"
  [ -n "$pid" ] && kill "$pid" 2>/dev/null
}

Endpoints

Method	Path	Purpose
POST	/v1/chat/completions	OpenAI chat completions (streaming + tools)
GET	/v1/models	models available to your account
GET	/usage	your rolling 5h / 7d usage percentages
GET	/health	liveness probe

Configuration (environment variables)

Var	Default	Meaning
COOKIE_FROM	–	firefox (auto-find profile) or a path to a profile dir / cookies.sqlite
CLAUDE_SESSION_KEY	–	sessionKey cookie value (manual alternative to COOKIE_FROM)
CLAUDE_COOKIE	–	full Cookie: header (manual; carries cf_clearance)
IMPERSONATE	firefox if COOKIE_FROM else chrome	curl_cffi TLS fingerprint
DEFAULT_MODEL	claude-sonnet-4-5-20250929	fallback slug when none matches
EFFORT	low	thinking effort: low / medium / high / xhigh / max
WEB_SEARCH	true	enable claude.ai's native web search
API_KEY	–	if set, require Authorization: Bearer <key>
HOST / PORT	127.0.0.1 / 8080	listen address
PROXY	–	upstream proxy URL for the claude.ai requests
DELETE_AFTER	true	delete reused threads on exit / when idle
SESSION_TTL	1800	seconds a reused thread is kept before it's dropped + deleted
MAX_SESSIONS	40	max reused threads kept at once
MAX_PROMPT_CHARS	120000	trim old turns past this (full-send fallback only)
TOOL_RESULT_CAP	6000	per-tool-result character cap

How it works

claude.ai's web completion endpoint takes a single prompt per turn and has no native function-calling. This proxy:

1. Reads your browser cookies and impersonates the browser's TLS fingerprint so Cloudflare is satisfied.
2. For the first turn of a session it sends the system prompt + tool schemas; after that it reuses the same claude.ai thread and sends only each new turn (tool results / follow-ups), letting Claude hold the context server-side.
3. Injects tool schemas + instructions so Claude emits a <tool_calls> JSON block, then parses that back into OpenAI tool_calls.

Limitations

• Prompt-level tool-calling is less reliable than a native function-calling API. It works well with Sonnet but can occasionally need a retry.
• No token counts (claude.ai doesn't send any) — usage is the 5h / 7d percentages only.
• Usage limits are the real ceiling. claude.ai meters a rolling 5-hour and 7-day window (not tokens), and they are small on the free plan. Agentic use makes many calls per task, so the 5h window can be exhausted quickly — then you are rate-limited until it refills. Watch it with GET /usage.
• Model access depends on your plan. Sonnet is available on the free plan; Opus and other top models require a paid plan and otherwise return 403 "upgrade your plan".
• Firefox only for automatic cookie reading (its cookies are unencrypted on disk on Linux).
• Unofficial; see the disclaimer above.