@iznogoudd/pi-devin

Devin Local models in Pi. Auth comes from the local Devin CLI / Devin Desktop session and the model catalog from the live GetCliModelConfigs RPC.

Fork of sting8k/pi-devin, itself a fork of kashyab12/pi-devin (npm pi-devin). Install only one of them: they all register the same devin provider.

How it works

Devin credentials are the only stored secret — every model call is plain HTTP, no OpenAI-compat layer.

  ~/.local/share/devin/credentials.toml        (windsurf_api_key — written by `devin auth login`)
        │
        │  /login devin  (only when the file is missing: spawn `devin auth login`,
        │                or reuse a signed-in Devin Desktop token)
        ▼
  ┌─────────────────────────────────────────────────────────────┐
  │ Model request  (per chat message)                           │
  │                                                             │
  │  POST {api_server_url}/exa.api_server_pb.ApiServerService/  │
  │       GetChatMessage                                        │
  │   • Metadata.api_key (field 3)  ← the windsurf_api_key      │
  │   • Metadata.user_jwt (field 21) ← minted via GetUserJwt    │
  │     (api key → GetUserJwt → 15-min JWT; cloud models work   │
  │      with the api key alone, JWT carries plan/entitlement)  │
  │   • system prompt (2), history (3), tools (10), config (8), │
  │     chat_model_uid (21) e.g. "swe-1-7-medium"               │
  │   • Connect RPC + protobuf (application/connect+proto,      │
  │     gzip framing) — NOT an OpenAI-compatible endpoint       │
  └─────────────────────────────────────────────────────────────┘
        ▲
        │  model list: GetCliModelConfigs over HTTP (ide=windsurf),
        │  `devin models list` only as fallback
        └─ never spawns the CLI during normal chat

Why this exists

Older packages treated Devin as Cascade cloud chat. Models like Sol High, Opus 5, and Fable 5 then failed with:

This model is only in Devin Local.

Those models are only reachable through the local Devin surface. This package reuses that local credential, then streams completions into Pi so Pi's tools, sessions, and UI stay in charge — it is not an ACP integration and does not use Zed.

Requirements

• Devin CLI installed (or bundled with Devin Desktop)
• Signed in: devin auth login, or a signed-in Devin Desktop on this machine
• Pi Coding Agent — @earendil-works/pi-ai and @earendil-works/pi-coding-agent are peer deps

Install

pi install npm:@iznogoudd/pi-devin

Local checkout:

pi install ~/Developers/pi-devin

Restart Pi or run /reload.

Usage

/login devin
/model devin/swe-2
/model devin/claude-opus-5
/model devin/gpt-5.6-sol

Models keep their family id; thinking levels belong to pi and each level is resolved to the matching Devin variant. For SWE-2:

| pi thinking level | model uid sent | |---|---| | medium | swe-2-medium | | high (default) | swe-2-high | | max | swe-2-max |

Use /thinking or shift+tab to change the level; levels a family does not ship are hidden, and Ctrl+S in /thinking saves the startup default. The same applies to every other family (devin/kimi-k3, devin/grok-4.6, …).

/login devin seeds ~/.local/share/devin/credentials.toml from a Devin Desktop sign-in you already have, and otherwise runs devin auth login.

Thinking: the server streams a summary of the model's reasoning; the full trace stays inside the sealed signature and never leaves the server. Pi keeps that summary together with its signature and replays both on the next request, exactly like the Devin CLI, so the model keeps its own prior reasoning across tool calls and turns.

Latency

The provider pre-warms the user JWT and TLS connection at startup and session start, so the first message skips the ~0.85s of cold-start cost. Each turn sends exactly one GetChatMessage request — hedged duplicate requests were removed once Devin introduced rate limiting.

Commands:

• /devin-status — CLI path, version, auth
• /devin-refresh — reload the model catalog

License

MIT. Unofficial. Not affiliated with Cognition.