pi-commandcode-provider

A pi extension that adds Command Code as a model provider — DeepSeek V4 Pro/Flash, Kimi K3, GLM 5.2, MiniMax M3, Qwen 3.7 Max, Inkling, and the rest of the open-weight roster, all routable from inside pi.

Unofficial. Reverse-engineered from the command-code npm CLI (currently verified against v0.52.1). Not affiliated with or endorsed by Command Code / Langbase. Schema is undocumented and can change without notice.

Why this exists

Command Code's /provider/v1/messages and /provider/v1/chat/completions generation endpoints require the Provider plan or higher. The Go plan ($1/month with $10 of credits and per-model multipliers, e.g. ~$40 of DeepSeek V4 Pro) doesn't expose them. (The GET /provider/v1/models list endpoint is readable on Go — handy for discovering ids — but generation isn't.)

This extension instead talks to /alpha/generate, the endpoint the Command Code CLI (cmd) uses for every model call. It's not plan-gated, so any account that can run cmd can use it. The trade-off is that the request/response shape is undocumented and Vercel-AI-SDK-flavored, not OpenAI- or Anthropic-shaped — hence the extension.

Configure

Set your Command Code API key as an env var. You can mint one at commandcode.ai/settings (it looks like user_…):

export COMMANDCODE_API_KEY=user_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

Verify the key is live:

curl -sS https://api.commandcode.ai/alpha/whoami \
  -H "Authorization: Bearer $COMMANDCODE_API_KEY"

Optional: set CMD_ZDR=1 to send Command Code's strict zero-data-retention header. Models without a ZDR-capable upstream will fail with 422 instead of falling back.

Install

pi install git:github.com/safzanpirani/pi-commandcode-provider

Or try it for one run without installing:

pi -e git:github.com/safzanpirani/pi-commandcode-provider

Use

pi --model "commandcode/deepseek/deepseek-v4-pro" -p "what is 17 * 23?"
pi --model "commandcode/moonshotai/Kimi-K2.6" -p "draft a haiku about caching"

pi --list-models | grep commandcode shows everything registered.

Models

Pulled from GET /provider/v1/models and the CLI registry, then registered with the canonical id the gateway expects. All support text; models marked by the CLI as multimodal also accept image input. This is the open-weight roster; proprietary frontier models are omitted by default (see below).

Model id	Display name	Context	Max out
deepseek/deepseek-v4-pro	DeepSeek V4 Pro	1M	128K
deepseek/deepseek-v4-flash	DeepSeek V4 Flash	1M	128K
moonshotai/Kimi-K3	Kimi K3	1M	64K
moonshotai/Kimi-K2.7-Code	Kimi K2.7 Code	256K	64K
moonshotai/Kimi-K2.7-Code-Highspeed	Kimi K2.7 Code HighSpeed	262K	64K
moonshotai/Kimi-K2.6	Kimi K2.6	256K	64K
moonshotai/Kimi-K2.5	Kimi K2.5	256K	64K
zai-org/GLM-5.2	GLM 5.2	1M	128K
zai-org/GLM-5.2-Fast	GLM 5.2 Fast	1M	64K
zai-org/GLM-5.1	GLM 5.1	200K	32K
zai-org/GLM-5	GLM 5	200K	32K
MiniMaxAI/MiniMax-M3	MiniMax M3	1M	128K
MiniMaxAI/MiniMax-M2.7	MiniMax M2.7	200K	64K
MiniMaxAI/MiniMax-M2.5	MiniMax M2.5	200K	64K
xiaomi/mimo-v2.5-pro	MiMo V2.5 Pro	1M	128K
xiaomi/mimo-v2.5	MiMo V2.5	1M	128K
Qwen/Qwen3.7-Max	Qwen 3.7 Max	1M	128K
Qwen/Qwen3.7-Plus	Qwen 3.7 Plus	1M	128K
Qwen/Qwen3.6-Max-Preview	Qwen 3.6 Max Preview	200K	32K
Qwen/Qwen3.6-Plus	Qwen 3.6 Plus	200K	32K
stepfun/Step-3.7-Flash	Step 3.7 Flash	256K	64K
stepfun/Step-3.5-Flash	Step 3.5 Flash	1M	128K
tencent/Hy3	Tencent Hy3	262K	64K
nvidia/nemotron-3-ultra-550b-a55b	Nemotron 3 Ultra	1M	128K
thinkingmachines/inkling	Inkling	256K	64K

Proprietary models (omitted by default). Command Code also serves claude-opus-4-8, claude-opus-4-7, claude-sonnet-4-6, claude-fable-5, claude-haiku-4-5-20251001, gpt-5.5, gpt-5.4, gpt-5.4-mini, gpt-5.3-codex, google/gemini-3.5-flash, and google/gemini-3.1-flash-lite over the same envelope. They bill real plan credits and are usually cheaper elsewhere, so they aren't registered here.

To add any model, append a ModelDef to the MODELS array in index.ts — every id in the gateway's GET /provider/v1/models list works.

How it works

• POST https://api.commandcode.ai/alpha/generate with Authorization: Bearer <key>
• Body is a wrapped envelope: { config, memory, taste, skills, permissionMode, params } where params carries the actual { model, system, messages, tools, max_tokens, stream }
• config, memory, taste, and skills are static neutral defaults in index.ts. A fork that wants Command Code's workspace/taste context can replace staticConfig() and those sidecar fields without touching message or stream parsing.
• Messages use the Vercel AI SDK ModelMessage schema — tool-call parts on assistant messages, role: "tool" + tool-result parts for tool outputs (not Anthropic content blocks, not OpenAI tool messages)
• Response is NDJSON (newline-delimited JSON), not SSE — events like reasoning-start/delta/end, text-start/delta/end, tool-input-start/delta/end, finish-step
• The extension translates pi's native message format ↔ that wire shape and maps stream events onto pi's thinking_* / text_* / toolcall_* events

For the full, field-by-field contract, see the docs below.

Docs

Reverse-engineered reference for the undocumented endpoint this extension speaks — useful if you're forking, extending, or the schema drifts and something starts 400ing:

• docs/wire-protocol.md — the complete /alpha/generate request/response contract: config envelope, Vercel-AI-SDK ModelMessage schema, tool format, every NDJSON event type, usage shape.
• docs/caching.md — prompt caching behavior + measured numbers (~36× cheaper on a warm prefix), and why the extension subtracts cached tokens from input.
• docs/troubleshooting.md — the 401 / 403 / 400 error modes, what each means, how to fix, and how to re-derive the protocol from the CLI bundle when it changes.

Caveats

• Undocumented endpoint. /alpha/generate isn't a published API surface. Command Code can change the schema at any time. If they tighten the request shape, expect 400 errors until the extension is updated. For request-schema 400s, first compare the outgoing body in index.ts against the current CLI envelope: config.{workingDir, date, environment, structure, isGitRepo, currentBranch, mainBranch, gitStatus, recentCommits}, memory: string, taste, skills, permissionMode, and params.{model, system, messages, tools, max_tokens, stream}.
• Tier policy. This extension uses an endpoint the official CLI uses, on credentials your plan grants. Command Code may consider that fair game or may not; their published "API access" feature still requires the Pro plan. If your account gets flagged or rate-limited, that's the realistic downside.
• Vision depends on the model. Kimi K3/K2.x, MiniMax M3, Qwen 3.7 Plus, Step 3.7 Flash, and Inkling currently declare image input.
• Cost is reported as $0 by pi. The Go plan applies per-model multipliers (e.g. $10 credits ↔ ~$40 of DeepSeek usage) that don't map cleanly to per-token pricing. Real balance: curl -H "Authorization: Bearer $COMMANDCODE_API_KEY" https://api.commandcode.ai/alpha/billing/credits.
• Reasoning is heavy. DeepSeek V4 Pro and Qwen 3.7 Max emit a lot of reasoning tokens by default. A short answer can cost 10–30× the input. Budget accordingly on the Go plan.

Development

git clone https://github.com/safzanpirani/pi-commandcode-provider
cd pi-commandcode-provider

# Test it without installing
COMMANDCODE_API_KEY=user_... pi -e . -p "say pong"

# Type-check
npm install
npm run check
npm test

Pi loads .ts files directly, so there's no build step.

License

MIT — see LICENSE.