pi-bai

A standalone pi extension for B.AI — a unified LLM gateway whose API is compatible with the OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages protocols.

• Chat is routed through the OpenAI Responses API (POST /v1/responses), which exposes native reasoning controls (reasoning.effort / reasoning.summary). The extension delegates to pi-ai's built-in openai-responses core, so streaming, tool calls, thinking blocks, and reasoning all work out of the box.
• Image generation uses the OpenAI-compatible POST /v1/images/generations (gpt-image-2, grok-imagine-image-2.0), saved under .pi/generated-images/.

Docs: https://docs.b.ai/llmservice/api/

Authentication

Store the API key in Pi with:

/login bai

Pi saves it in ~/.pi/agent/auth.json. BAI_API_KEY remains supported as a fallback for existing setups.

# macOS / Linux / WSL
export BAI_API_KEY="sk-..."

# Windows PowerShell
$env:BAI_API_KEY = "sk-..."

Install like any pi extension:

npm install pi-bai

Then restart pi so the bai provider is loaded. The model catalog is fetched automatically from GET https://api.b.ai/v1/models on startup and overrides the built-in seed list.

How it works

B.AI exposes three protocol-compatible endpoints under the production base URL https://api.b.ai/v1:

Endpoint	Protocol	Used here
/v1/chat/completions	OpenAI Chat Completions	✅ chat for chat-only models
/v1/responses	OpenAI Responses	✅ chat for reasoning-capable models
/v1/messages	Anthropic Messages	—
/v1/images/generations	OpenAI Images	✅ image generation

• Auth — BAI_API_KEY sent as Authorization: Bearer <key>.
• Reasoning — chat models flagged reasoning: true (GPT-5.x, Claude Opus 5, Gemini 3.6, DeepSeek v4, GLM-5.3, Kimi, Grok, MiMo, MiniMax, Hunyuan) get a thinkingLevelMap so pi's thinking-level selector maps to B.AI's none / low / medium / high / xhigh / max efforts. By default reasoning is none; selecting a level in pi sends reasoning: { effort, summary }.
• Per-model endpoint selection — B.AI rejects some models on /v1/responses with model_not_supported_on_endpoint. The chat/completions-only models we know about (deepseek-v4-flash, deepseek-v4-flash-vision-exp, hy3, mimo-v2.5, glm-5.3-flash) are flagged baiChatOnly: true and routed straight to the OpenAI Chat Completions core. Every other model is tried on the Responses API first (for native reasoning effort / summary), and if B.AI rejects it as /v1/responses-only, the request is transparently retried on /v1/chat/completions (which supports all models; reasoning params are dropped on the retry since those models reason via native tokens). So you don't need to enumerate chat-only models — the fallback covers any we haven't listed. Chat-only models still reason natively (the core surfaces their reasoning_content thinking tokens); they just don't expose B.AI's Responses-API effort/summary controls.
• System role — B.AI's upstream rejects the OpenAI developer message role (HTTP 400, code 1214, "角色信息不正确"). Every chat/completions request is forced to use role: "system" for the system prompt (via compat.supportsDeveloperRole: false on the resolved model), so reasoning-capable chat-only models that hit the Responses→Chat-Completions fallback also work.
• Model catalog — GET /v1/models returns an OpenAI-compatible { object, success, data: [{ id, object, created }] }. refreshModels fetches it on startup / on demand, publishes the result, and always merges the image models (which are not part of the /v1/models catalog).
• Free tier & access status — B.AI does not advertise free/premium in /v1/models, and pricing changes over time (free ⇄ charge). The signal is the API response, and the extension learns it automatically:
  • 200 from /v1/chat/completions → FREE
  • 403 access_denied ("Deposit required to unlock premium models") → CHARGE
  • 400 insufficient_user_quota → QUOTA Every chat stream is observed, so the moment you use a model that flipped free→charge (or charge→free), the /bai-models badge corrects itself. Unknown models fall back to the seeded free list below. The currently-free models (confirmed by probing the catalog) are: deepseek-v4-flash, deepseek-v4-flash-vision-exp, hy3, mimo-v2.5, glm-5.3-flash, qwen3.8-flash. For a full refresh on demand, run /bai-free (it probes every model against /v1/chat/completions and updates all badges).
• Image generation — dispatched from streamBai via the baiImageModel flag, the same pattern pi-cloudflare-workers-ai uses (pi's extension API has no separate image-registration surface). The exact image endpoint is assumed OpenAI-compatible; change IMAGE_GEN_PATH in extensions/bai.ts if your plan differs.

Models

The seed list covers B.AI's language-model families and image models. After the first refresh, the live GET /v1/models catalog replaces the chat seed (image models are always merged back in), so you always see exactly the models enabled for your API key.

Capabilities. B.AI's GET /v1/models returns only { id, object, created } — it does not expose context windows, max output, or modalities. The extension therefore ships a per-model capability table (MODEL_CAPS in extensions/bai.ts) sourced from each model's public specifications, with a family-based fallback (familyCaps) for any model the live catalog adds later. The /bai-models command shows Context, Max Out, and Modalities columns so you can see each model's real limits at a glance:

Family	Context	Max out	Input
gpt-5.*	400K	16K–128K	text + image
claude-*	200K	64K	text + image
gemini-*	1M	64K	text + image
deepseek-v4*	128K	32K	text (+ image for vision-exp/pro)
glm-5.*	128K	32K	text + image
kimi-*, qwen3.8*, minimax-*, hy3, mimo-*	256K	32K	text

These are best-effort figures from public model specs. If B.AI's actual limits differ, edit MODEL_CAPS (or report it) — they are not fetched from the API.

Commands

Commands

/bai-models

Lists the B.AI models currently available to pi. By default it shows only the confirmed FREE chat models and hides charged / quota models (image models are listed under all, since their pricing isn't exposed by B.AI):

/bai-models

Pass all to list every model, including charged, quota-limited, and image models:

/bai-models all

/bai-docs

Opens the B.AI API reference in your browser.

/bai-docs

/bai-free

Probes every B.AI model against /v1/chat/completions and refreshes the FREE / CHARGE / QUOTA status for all of them (pricing changes over time, so this re-syncs the badges). It reports the counts and the model ids in each tier.

/bai-free

The status is also learned automatically from your normal chat usage, so you only need /bai-free when you want an immediate full refresh.

Notes

• Generated images are saved under .pi/generated-images/ and reported as a clickable link (OSC 8 hyperlink); the TUI also renders an inline preview.
• If BAI_API_KEY is unset, pi marks the bai provider as unconfigured and the seed list is shown until a key is provided and a refresh succeeds.
• The image-generation endpoint shape was inferred from B.AI's "GPT-Image-2 is an OpenAI image generation … model" description (docs image pages were rate-limited at implementation time). Verify IMAGE_GEN_PATH against the live docs if needed.