Crystal Code Defaults

Default provider and model definitions for Crystal Code.

This repository holds the reference providers.json catalog: named endpoints, protocols, model ids, and per-model metadata such as context windows, output limits, thinking efforts, and image input. Copy it into your Crystal home to extend the small starter set built into Crystal Code with the full catalog.

Get started · Catalog · How the file is used · File format · Maintenance

📦 What's in this repository

File	Purpose
providers.json	The default provider and model catalog, generated by build_providers.py
providers/	One JSON file per provider. Edit these, then regenerate the catalog
build_providers.py	Merges providers/*.json into providers.json

🚀 Get started

📥 1. Download the catalog

Crystal Code reads providers.json from its home directory: ~/.crystal/providers.json, or the directory named by CRYSTAL_HOME when that variable is set. On Windows the default home is %USERPROFILE%\.crystal.

Linux or macOS

mkdir -p ~/.crystal
curl --fail --location --show-error \
  https://raw.githubusercontent.com/YELANDAOKONG/CrystalCode.Defaults/master/providers.json \
  --output ~/.crystal/providers.json

Windows PowerShell

$crystalHome = "$env:USERPROFILE\.crystal"
New-Item -ItemType Directory -Force $crystalHome | Out-Null
Invoke-WebRequest `
  -Uri https://raw.githubusercontent.com/YELANDAOKONG/CrystalCode.Defaults/master/providers.json `
  -OutFile "$crystalHome\providers.json"

The download replaces an existing providers.json. Back up or merge a file you already customized before you overwrite it.

🔑 2. Set provider keys

Keys never belong in this file. Write each one in ~/.crystal/credentials.json under the provider name, or set <PROVIDER>_API_KEY for the process that starts the app. A provider entry can also name an environment variable with apiKeyEnvironment, as bigmodel does with ZHIPU_API_KEY. See Credentials in the user guide for the full lookup order.

🏁 3. Restart Crystal Code

Provider definitions are read at startup. Restart the app after the file changes, then select a model with /model, or start with --provider and --model.

🗂️ Catalog

Provider	Protocols	Base URI
deepseek	deepseek	https://api.deepseek.com/
openai	openai	https://api.openai.com/v1/
anthropic	anthropic	https://api.anthropic.com/v1/
gemini	gemini	https://generativelanguage.googleapis.com/v1beta/
moonshot	openai	https://api.moonshot.cn/v1/
siliconflow	openai	https://api.siliconflow.cn/v1/
bigmodel	openai	https://open.bigmodel.cn/api/paas/v4/
openrouter	openai	https://openrouter.ai/api/v1/
requesty	openai	https://router.requesty.ai/v1/
opencode-zen	responses, anthropic, gemini, openai	https://opencode.ai/zen/v1/
opencode-go	responses, anthropic, openai	https://opencode.ai/zen/go/v1/
commandcode	openai, anthropic	https://api.commandcode.ai/provider/v1/
bigmodel-intl	openai	https://api.z.ai/api/paas/v4/
moonshot-intl	openai	https://api.moonshot.ai/v1/
opencode-free	openai	https://opencode.ai/zen/v1/
openrouter-free	openai	https://openrouter.ai/api/v1/
siliconflow-intl	openai	https://api.siliconflow.com/v1/
meta	openai	https://api.meta.ai/v1/
mistral	openai	https://api.mistral.ai/v1/
ollama-cloud	openai	https://ollama.com/v1/
stepfun	openai	https://api.stepfun.com/v1/
stepfun-intl	openai	https://api.stepfun.ai/v1/
together	openai	https://api.together.ai/v1/
xai	openai	https://api.x.ai/v1/
cerebras	openai	https://api.cerebras.ai/v1/
deepinfra	openai	https://api.deepinfra.com/v1/openai/
fireworks	openai	https://api.fireworks.ai/inference/v1/
groq	openai	https://api.groq.com/openai/v1/
huggingface	openai	https://router.huggingface.co/v1/
minimax	openai	https://api.minimaxi.com/v1/
minimax-intl	openai	https://api.minimax.io/v1/
nebius	openai	https://api.tokenfactory.nebius.com/v1/
novita	openai	https://api.novita.ai/openai/v1/
nvidia	openai	https://integrate.api.nvidia.com/v1/
qianfan	openai	https://qianfan.baidubce.com/v2/
qwen	openai	https://dashscope.aliyuncs.com/compatible-mode/v1/
qwen-intl	openai	https://maas.qwencloudapi.com/compatible-mode/v1/
sambanova	openai	https://api.sambanova.ai/v1/
tencent	openai	https://tokenhub.tencentcloudmaas.com/v1/
vercel	openai	https://ai-gateway.vercel.sh/v1/
volcengine	openai	https://ark.cn-beijing.volces.com/api/v3/
xiaomi	openai	https://api.xiaomimimo.com/v1/

A provider name can map to an array of entries when one service exposes several protocols or endpoints; opencode-zen, opencode-go, and commandcode use this form, and each model is routed to the protocol that owns it. moonshot-intl, siliconflow-intl, bigmodel-intl, qwen-intl, and minimax-intl are separate accounts from the domestic entries, so they do not reuse those API keys. openrouter-free and opencode-free still need an API key; the models themselves are the no-charge routes.

qwen is Alibaba Cloud Model Studio in China and qwen-intl is QwenCloud; minimax and minimax-intl use the China and international hosts. volcengine is ByteDance's Ark platform (Doubao), qianfan is Baidu's ERNIE platform, tencent is Tencent Cloud TokenHub (Hy), and xiaomi is the MiMo open platform. tencent defaults to the Guangzhou endpoint; use https://tokenhub-intl.tencentcloudmaas.com/v1/ for Singapore or https://tokenhub-us.tencentcloudmaas.com/v1/ for Silicon Valley. huggingface routes through Inference Providers, nvidia through the hosted NIM API, and vercel through AI Gateway.

Some large services are intentionally absent. Cohere's compatibility API rejects the stream_options field Crystal Code sends while streaming; Perplexity retired its Chat Completions surface in favor of the Agent API; GitHub Models was retired on 30 July 2026; Azure OpenAI, Amazon Bedrock, and Google Vertex AI need per-account or per-region base URIs; Cloudflare Workers AI embeds the account id in the URI. Other providers not covered in this pass include iFlytek Spark, ModelScope, Meituan LongCat, Hyperbolic, Featherless, Nscale, OVHcloud, Scaleway, Baseten, Modal, Replicate, Upstage, Sarvam, and Naver.

openai means OpenAI-compatible Chat Completions and responses is the OpenAI Responses API. deepseek, anthropic, gemini, and ollama are native protocols.

⚙️ How the file is used

• Without this file, Crystal Code falls back to a small starter catalog: deepseek, openai, gemini, and ollama. The anthropic protocol is built in, but has no starter entry.
• A providers.json in the Crystal home supplies the operator catalog and overlays the built-in one.
• Models that are not listed in the effective catalog cannot be selected. There is no global context window.
• A model may declare selectable context budgets with contextWindows; /context cycles them, and compaction and the status bar use the selected budget.
• When providers.json exists, the legacy providers object in config.json is ignored.

📄 File format

Top-level keys are provider names. Each value is one provider entry or an array of entries, in the same shape Crystal Code reads and writes.

Provider fields:

Field	Meaning
protocol	deepseek, openai, responses, anthropic, gemini, or ollama
baseUri	Absolute API base URI; the adapter appends the protocol's chat path
models	Table of selectable model ids
replayReasoningContent	Replay provider reasoning content (DeepSeek always does this)
tokenLimit	Chat Completions output field: max_tokens or max_completion_tokens
options	Optional object of protocol-specific settings; unknown names, or names that do not apply to the protocol, fail loading. temperature, topP, and maxTokens stay on the model
apiKeyEnvironment	Environment variable name checked first for this provider
apiKey	Literal, {env:NAME}, or {file:path}. Keep secrets out of this file
requiresApiKey	Whether a missing key is an error (default true, except ollama)
organization / project	Optional OpenAI organization and project

Model fields:

Field	Meaning
contextWindow	Required. Positive maximum token window; upper bound for selectable budgets, and the sole budget when contextWindows is omitted
contextWindows	Optional distinct positive token budgets no greater than contextWindow, for example [272000, 1000000]. /context cycles them; the lowest is the initial budget, and compaction and the status bar use the selected budget
maxTokens	Optional positive output-token cap
thinking	Whether the model accepts reasoning hints
thinkingEfforts	Portable names minimal, low, medium, high, and maximum (max is stored as maximum), or another effort token the endpoint accepts. A token starts with a letter and then uses only letters, digits, and hyphens
thinkingCanDisable	Whether /thinking off is accepted (default true). Models that always reason, such as Gemini 3, set this to false
imageInput	Whether the model may receive images (default false)
audioInput	Declared audio input (default false). Stored and visible to plugins; the host does not accept or send audio
videoInput	Declared video input (default false). Stored and visible to plugins; the host does not accept or send video
temperature	Optional, 0 to 2
topP	Optional, 0 to 1

The user guide covers protocols, examples, and credentials in full.

🤝 Maintenance

Maintainers can write a commit message and commit the current change with commit.sh. It runs crystal run in this repository. Extra arguments are appended to the task. The run skips external tools and the directory trust check, and it does not record the directory as trusted.

Edit the provider file under providers/ when a provider adds, retires, or changes models or model metadata. The file name, without .json, is the provider id. The file is one provider entry, or an array of entries when one service exposes several protocols.

Regenerate the catalog with build_providers.py. It overwrites providers.json and keeps two-space indentation. Provider order follows the current catalog; a provider file that is not already listed is appended in alphabetical order.

python3 build_providers.py

Crystal Code does not rewrite this file.

📄 License

MIT