Blueai Models

Configure and manage AI models from BlueAI unified proxy service for OpenClaw. Use when: (1) adding new models to openclaw.json, (2) choosing the right model...

Security Scan

The skill mostly does what it claims (add/test BlueAI models by editing ~/.openclaw/openclaw.json and calling a BlueAI relay), but the test script can send whatever API key it finds to a single default relay URL (potentially leaking non-BlueAI keys) and the declared environment/credential requirements don't match the code.

Purpose & Capability

Name/description match the included scripts and docs: the skill adds models to openclaw.json and tests connectivity against a BlueAI relay. Modifying ~/.openclaw/openclaw.json is expected for this purpose. However, some implementation choices (see below) are surprising for a 'test all configured models' operation.

Instruction Scope

SKILL.md instructs running the included scripts to add and test models which is consistent, but the test script's --all-configured mode does not use each provider's configured baseUrl; it uses a single base URL argument (defaulting to the BlueAI relay). That means testing 'all configured' will contact the relay for every model and will present whichever API key the script finds — not necessarily the key intended for that provider. This deviates from reasonable scope and can cause unintended key disclosure.

Install Mechanism

Instruction-only with lightweight Python helper scripts; no installer, downloads, or extracted archives. No suspicious install behavior.

Credentials

Metadata declares no required env vars, but test_model.py searches OPENAI_API_KEY and BLUEAI_API_KEY and add_model/test scripts read/write ~/.openclaw/openclaw.json to find/store apiKey values. The code can take a found API key and send it to the default relay — this is proportionate if the key is a BlueAI key, but dangerous if it's an unrelated provider key (e.g., a personal OpenAI key) because the script may transmit it to the relay without making that explicit.

Persistence & Privilege

add_model.py writes to the user's ~/.openclaw/openclaw.json to add providers/models (expected behavior). always:false and no global config modifications beyond the openclaw.json file. This is normal for a configuration helper, but users should be aware it will modify their config file.

What to consider before installing

This package largely does what it promises (editing ~/.openclaw/openclaw.json and calling a BlueAI relay), but there are two important gotchas you should understand before running it: 1) test_model.py will look for API keys in environment variables (OPENAI_API_KEY, BLUEAI_API_KEY) and inside your ~/.openclaw/openclaw.json and will use the first key it finds. If you run python3 scripts/test_model.py --all-configured (the undocumented risky case), the script will call a single base URL (default: https://bmc-llm-relay.bluemediagroup.cn/v1) for every configured model rather than each provider's configured endpoint — this can cause your unrelated provider keys (for example your OpenAI key) to be sent to the BlueAI relay. Do NOT run --all-configured unless you understand which key will be used. 2) The SKILL metadata declares no required env vars but the code does read env vars and the openclaw.json for apiKey values. Inspect the scripts yourself (they are small and included) before running. Prefer running targeted commands (test a single model and pass --api-key explicitly) or use a throwaway/test API key when exercising the scripts. Back up ~/.openclaw/openclaw.json before running add_model.py. If you want to proceed: review the two scripts, run test_model.py for single models with --api-key or ensure your config only contains keys intended for the BlueAI relay, and avoid --all-configured unless you update the script to use each provider's configured baseUrl and to avoid using unrelated keys.

BlueAI Models for OpenClaw

Quick Start

Add a model to OpenClaw:

python3 scripts/add_model.py gemini-2.5-flash --alias flash
python3 scripts/add_model.py claude-sonnet-4-6 --alias sonnet
openclaw gateway restart

Test connectivity:

python3 scripts/test_model.py gemini-2.5-flash
python3 scripts/test_model.py --all-configured

List available models:

python3 scripts/add_model.py --list

Image Generation

Gemini image models generate images via Chat Completions (/v1/chat/completions), not the Images API. Send a prompt as a normal message; the model returns base64-encoded images in Markdown.

# Add image models
python3 scripts/add_model.py gemini-3.1-flash-image-preview
python3 scripts/add_model.py gemini-3-pro-image-preview
openclaw gateway restart

# Test image generation
python3 scripts/test_model.py gemini-3.1-flash-image-preview --image-gen
python3 scripts/test_model.py gemini-3-pro-image-preview --image-gen --save ./test-output

Model	Speed	Quality	Edit Support	Best For
gemini-3.1-flash-image-preview	⚡ Fast	Good	❌	Quick prototypes, batch
gemini-3-pro-image-preview	Medium	⭐ Best	✅	High-quality creative
gemini-2.5-flash-image	⚡ Fast	Good	✅	Image editing

For detailed usage, prompt tips, Python examples, and edit workflows: read references/image-generation.md.

Endpoints

Type	Base URL	Note
Claude (Anthropic)	https://bmc-llm-relay.bluemediagroup.cn	No /v1
Everything else (OpenAI)	https://bmc-llm-relay.bluemediagroup.cn/v1	With /v1

Same API key works for all models.

Model Selection Quick Guide

Need	Model	Why
Cheapest + good	gemini-2.5-flash	$0.15/M in, 1M context
Best Chinese	DeepSeek-V3.2	Top Chinese quality, cheap
Vision + cheap	gpt-4o-mini or gemini-2.5-flash	Image input, low cost
Strong reasoning	o4-mini or DeepSeek-R1	CoT reasoning
Best overall	claude-opus-4-6-v1	128K output, Agent coding
Balanced	claude-sonnet-4-6	1/5 Opus price, most tasks
Code specialist	gpt-5.2-codex	128K output, code focused
Ultra-long context	xai.grok-4-fast-non-reasoning	2M tokens
Image gen (fast)	gemini-3.1-flash-image-preview	Chat-based, cheap
Image gen (quality)	gemini-3-pro-image-preview	Best Gemini quality
Image edit	gemini-2.5-flash-image	Send image + edit instruction

References

• Full model catalog: Read references/model-catalog.md for all 100+ models with specs
• OpenClaw config guide: Read references/openclaw-config.md for JSON structure and examples
• Model selection decision tree: Read references/model-selection.md for task-based recommendations
• Image generation guide: Read references/image-generation.md for Gemini image gen usage, prompts, and code examples

Key Rules

1. Claude models use api: "anthropic-messages", baseUrl without /v1
2. All other models use api: "openai-completions", baseUrl with /v1
3. DeepSeek/Qwen text models: set input: ["text"] only (no image)
4. MiniMax: must use OpenAI endpoint, does not support Claude endpoint
5. gemini-3-pro-preview deprecated 2026-03-26 → use gemini-3.1-pro-preview
6. Gemini image models use chat completions, not images API — output is base64 in Markdown
7. After config changes: openclaw gateway restart