Claude Code

Claude Code talks to devin-2api over the Anthropic Messages surface (POST /v1/messages). You need the proxy address and a downstream token from /web/tokens.html.

Setup

~/.claude/settings.json:

{
    "env": {
        "ANTHROPIC_BASE_URL": "http://127.0.0.1:8080",
        "ANTHROPIC_AUTH_TOKEN": "<downstream token>",
        "ANTHROPIC_MODEL": "swe-2-max",
        "ANTHROPIC_SMALL_FAST_MODEL": "swe-2-max",
        "CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY": "1",
        "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1",
        "CLAUDE_CODE_MAX_CONTEXT_TOKENS": "262000",
        "CLAUDE_CODE_AUTO_COMPACT_WINDOW": "230000"
    }
}

Replace 8080 with your server.listen port. Model name: swe-2-max, or any name you set in devin.aliases / the model registry.

Why the two window variables matter

swe-2-max doesn't carry a claude- prefix, so Claude Code falls back to its unknown-model context window — far below the real upstream limit of 262000. Without the declarations, auto-compact either fires too early (wasting the window) or sits above the true limit (requests die of prompt-too-long instead of compacting first). AUTO_COMPACT_WINDOW at 230000 leaves ~30k headroom for the compaction request's own instructions and summary.

The alternative: let Claude Code send a claude--prefixed name and map it back via devin.aliases — the window math then follows the claude model's profile.

Optional resilience variables

{
    "env": {
        "CLAUDE_CODE_MAX_RETRIES": "15",
        "CLAUDE_STREAM_FIRST_BYTE_TIMEOUT_MS": "300000"
    }
}

• CLAUDE_CODE_MAX_RETRIES — upstream rate-limit episodes can run for several minutes; Claude Code sleeps per the anthropic-ratelimit-unified-reset header and each retry spends one budget slot. A higher budget rides out longer episodes (non-first-party mode clamps at 15).
• CLAUDE_STREAM_FIRST_BYTE_TIMEOUT_MS — first-byte watchdog in milliseconds; upstream can think silently for a while, so 300000 keeps long prefills from tripping it.

WebSearch

Claude Code's WebSearch normally runs as a separate side request carrying a web_search server tool — devin-2api detects that shape and short-circuits it into a single managed search through the upstream search RPC, returning the result as a native server_tool_use block. No configuration needed.