Other Clients

Any client that speaks OpenAI Chat Completions, OpenAI Responses, or Anthropic Messages can use devin-2api. Two values are always needed: the proxy address (server.listen, examples here use http://127.0.0.1:8080) and a downstream token from /web/tokens.html.

Generic SDKs

• Anthropic Messages — point the SDK's base URL at http://127.0.0.1:8080 and the API key at your downstream token; requests land on POST /v1/messages.
• OpenAI — base URL http://127.0.0.1:8080/v1, API key = downstream token. chat.completions.create hits /v1/chat/completions; Responses-API clients hit /v1/responses.

Send the token as Authorization: Bearer <token> or X-Api-Key: <token>.

pi

~/.pi/agent/models.json:

{
    "providers": {
        "devin": {
            "baseUrl": "http://127.0.0.1:8080",
            "api": "anthropic-messages",
            "apiKey": "<downstream token>",
            "models": [
                {
                    "id": "swe-2-max",
                    "contextWindow": 262000,
                    "maxTokens": 32768,
                    "reasoning": true,
                    "input": ["text", "image"]
                }
            ]
        }
    }
}

Run with pi --provider devin --model swe-2-max, or pick the model interactively via /model. pi has no permission system — tools all execute by default; narrow with --tools read,grep,find,ls (allowlist), --no-builtin-tools, or --no-tools.

kimi-code

~/.kimi-code/config.toml:

default_model = "swe-2-max"

[providers.devin]
type = "anthropic"
base_url = "http://127.0.0.1:8080"
api_key = "<downstream token>"

[models."swe-2-max"]
provider = "devin"
model = "swe-2-max"
max_context_size = 262000
capabilities = ["thinking", "tool_use", "image_in"]

Unknown model names need an explicit capabilities list or tool calling never engages. Valid values: thinking / always_thinking / tool_use / image_in / video_in. type = "openai" (chat completions) and "openai_responses" also work. Permission modes and [[permission.rules]] are documented upstream — default_permission_mode accepts manual / yolo / auto.

Shared notes

• Context windows — declare the real 262000 wherever the client asks; a misdeclared window breaks auto-compaction timing (see each client's section).
• System-prompt fingerprints — upstream content policy can reject a client's identity prompt (permission_denied). devin-2api's sanitizer covers Claude Code's fingerprint; pi and kimi-code masquerade as Claude Code, so the same rules cover them.
• Tool-call pairing — upstream requires call→result adjacency; the proxy re-pairs automatically and clients don't notice.
• Prefer http://[::1]:<port> for the base URL when the service binds *: 127.0.0.1 can be silently shadowed by an IDE's IPv4 port forward (connect succeeds, zero bytes flow), and localhost depends on resolver order.