| name | kiln-cli |
| description | Use Kiln as a pure OpenAI-compatible CLI for local chat, translation, embeddings, OCR, model listing, and backend health checks. Trigger whenever the user asks to use a local model, translate text, generate embeddings, OCR an image, inspect available models, or diagnose the local API. |
| compatibility | Requires the `kiln` executable and an OpenAI-compatible backend reachable through `KILN_BASE_URL`. |
Kiln CLI
Kiln is only an API client. It does not start, stop, restart, unload, configure, or supervise the backend. Do not use launchd commands, kiln service, kiln serve, kiln unload, or kiln config; those are no longer CLI commands.
Keep the API key private. Never print it, copy it into an artifact, or put it in a shell command shown to the user. The local default key is local; use KILN_API_KEY or KILN_API_KEY_FILE when the backend expects another value.
Environment
export KILN_BASE_URL=http://127.0.0.1:8007
export KILN_API_KEY=local
export KILN_MODEL=ornith-ai--Ornith-1.5-35B-A3B-MLX-4bit
export KILN_OCR_MODEL=Unlimited-OCR-mxfp8
export KILN_EMBEDDING_MODEL=mlx-community--Qwen3-Embedding-4B-4bit-DWQ
export KILN_TRANSLATE_MODEL=Hy-MT2-1.8B-4bit
OCR and translation fall back to KILN_MODEL when their specific model variable is unset. Embeddings have no fallback and fail clearly when KILN_EMBEDDING_MODEL is unset.
rich is a direct CLI dependency declared in pyproject.toml; uv resolves and caches it on first run. TTY output gets Markdown and table rendering; pipes stay machine-readable. Use KILN_RICH=never to force plain output or KILN_RICH=always to force rich rendering.
Safe start
Run kiln doctor before an operation that depends on the backend when availability is uncertain. If it is down, report the endpoint and ask the user to repair or start the backend externally. Do not run backend recovery commands on Kiln's behalf.
The CLI retries transient 5xx and connection failures for 12 attempts, 3 seconds apart by default. Do not add a competing retry loop. Override with KILN_RETRIES and KILN_RETRY_DELAY only when needed.
Commands
Chat
kiln chat '解释一下什么是 RAG'
cat document.txt | kiln chat
It sends one user message to /v1/chat/completions using KILN_MODEL and prints plain assistant text. stdin is preferred for long or multiline input.
Translation
kiln translate '今天天气不错。'
kiln translate --lang 中文 'The weather is nice today.'
The default target is English; --lang or -l selects another target language. The CLI adds a translation-only instruction because Hy-MT2 otherwise may behave like a general chat model. Override it with KILN_TRANSLATE_INSTRUCTION.
Embeddings
kiln embed '这段文字用于向量检索'
cat document.txt | kiln embed
It sends to /v1/embeddings using KILN_EMBEDDING_MODEL and prints raw OpenAI-compatible JSON. Do not paste large vectors into a normal response; report count and dimension unless raw JSON was requested.
OCR
kiln ocr ./invoice.png
kiln ocr '只输出图片中的文字' ./invoice.png
kiln ocr https://example.com/image.png
OCR sends a multimodal request to /v1/chat/completions using KILN_OCR_MODEL. Local images become data URLs; HTTP(S) image URLs are passed through. The default instruction is 请识别图片中的全部文字并原样输出; override it with KILN_OCR_INSTRUCTION or a positional instruction.
OCR prints the model response. It does not run PaddleOCR and does not create Markdown, JSON, DOCX, layout PNG, or an output directory. PDF and office containers are rejected; convert them to images first.
Models and health
kiln models
kiln doctor
models lists /v1/models ids. In a TTY it uses a Rich table when available; pipes receive one id per line. doctor prints /health JSON, or a Rich table in a TTY. Neither command exposes the API key.
Logs
Kiln does not own backend logs. kiln logs only works when KILN_LOG points to an existing backend log file:
KILN_LOG=/path/to/backend.log kiln logs -f
Failure handling
- Missing image or invalid syntax: show the corrected
kiln command; do not guess a path.
- Backend unavailable: report the configured endpoint and ask for external backend recovery.
- Missing embedding model: tell the user to set
KILN_EMBEDDING_MODEL; never fall back to a chat model.
- OCR failure: preserve the input and report the backend error. There is no Kiln-managed output directory.
- API-key errors: do not inspect or repeat the key; ask the user to repair their local environment.
Final report
→ Result. What completed or failed.
→ Backend. Endpoint and model only if relevant, never the secret.
→ Output. Plain text, JSON path, or OCR result location if applicable.
