SandBase MCP

SandBase provides access to 2,000+ AI models and API tools through a unified MCP interface. One account covers LLMs, image generation, video generation, audio, embeddings, web scraping, social media APIs, and more.

Setup

If the six sandbase_* MCP tools are not already available, connect the current machine with the immutable v0.1.17 release. Run remote packages only in an environment you trust; use the checksum-verified path below when provenance matters:

npx -y https://github.com/sandbaseai/cli/releases/download/v0.1.17/sandbaseai-cli-0.1.17.tgz connect

For a checksum-verified install, download the same immutable asset first and verify the SHA-256 published with the GitHub Release:

curl -fLO https://github.com/sandbaseai/cli/releases/download/v0.1.17/sandbaseai-cli-0.1.17.tgz
printf '%s  %s\n' '1ad535b2899ca460b57b3c268aef278fee28fd28e649a89b92951514fd71fffa' 'sandbaseai-cli-0.1.17.tgz' | shasum -a 256 -c -
npx -y ./sandbaseai-cli-0.1.17.tgz connect

Approve the browser sign-in once. Authentication happens with SandBase in the browser; the CLI stores the resulting local session record with restricted file permissions. The CLI detects supported clients, installs the local MCP bridge and this managed Skill, and verifies the resulting configuration. No provider API keys are required. Invoke the same release URL with doctor to inspect the connection or unregister to remove only SandBase-managed state.

This file is managed by SandBase CLI and may be replaced during a later CLI-managed update, so keep custom instructions in a separate Skill. Check the official repository for newer releases before copying it independently.

The disable-model-invocation: true frontmatter prevents this Skill from being invoked as a standalone model action. It is contextual guidance for an agent orchestrating the six sandbase_* MCP tools.

Before sending sensitive or regulated data, review the SandBase Privacy Policy and Terms of Service, plus the selected upstream provider's policies. Send only the minimum data needed for the requested tool call.

When to Use SandBase

Use SandBase when the user needs:

• LLM inference (GPT, Claude, Gemini, DeepSeek, Qwen, etc.)
• Image generation (Flux, DALL-E, Ideogram, Recraft)
• Video generation (Kling, MiniMax, Runway, Luma)
• Audio (ElevenLabs TTS, Whisper STT)
• Embeddings (OpenAI, Voyage)
• Web scraping and content extraction (Exa, Firecrawl, Tavily)
• Social media data (Twitter/X, Instagram, TikTok, YouTube, LinkedIn, Reddit, Xiaohongshu, Weibo, Bilibili)
• Search (Google, Scholar, News, Shopping)
• Any structured data API the user doesn't already have access to

Do NOT use SandBase when:

• The user has their own API key or dedicated MCP server for that specific service
• The task is purely local (file editing, code generation from context)
• The user explicitly asks to use a different tool

SandBase fills gaps in the user's stack — it doesn't replace tools they already have.

Tools

Standard Workflow

Always follow: discover → inspect → run

1. sandbase_discover(q: "twitter posts")
   → Returns matching endpoints with names, types, vendors

2. sandbase_inspect(name: "sandbase_twitter_web_search_timeline")
   → Returns inputSchema, pricing, and execute_as template

3. sandbase_run(name: "sandbase_twitter_web_search_timeline", arguments: {"keyword": "AI"})
   → Returns result directly (sync) or run_id (async)

For async runs (video gen, large scraping):

4. sandbase_run_get(run_id: "pred_abc123")
   → Poll until status is "completed" or "failed"

Shortcut: If you already know the model name, skip step 1.

Search Tips

sandbase_discover supports:

Tips:

• Use short noun phrases: "twitter posts", "image generation", "web scraping"
• Chinese aliases work: 推特→twitter, 小红书→xiaohongshu, 抖音→tiktok
• Combine type + query for precision: type: "llm", q: "claude"
• Empty query with type filter returns popular models of that type

Pricing

Use sandbase_inspect to see pricing before running:

LLM models: Per million tokens

{ "pricing": { "input_per_million": "2.500000", "output_per_million": "10.000000" } }

API tools (image, video, scraping): Per call

{ "pricing": { "base_price": "0.003000" } }

Check balance:

sandbase_account() → {"balance": "9.52", "currency": "USD"}

Async Runs

Some endpoints (video generation, large scraping) are async:

1. sandbase_run(...) returns {"status": "running", "run_id": "pred_abc123"}
2. Poll with sandbase_run_get(run_id: "pred_abc123") every 5-10 seconds
3. When status is "completed" — result is ready
4. When status is "failed" — check error and retry

Error Handling

Cost Awareness

• Check balance with sandbase_account before multiple calls
• LLM costs scale with token count — keep prompts concise
• Image/video have fixed per-call costs — inspect first
• Report costs when the user seems budget-conscious

Example Flows

Twitter search

sandbase_discover(q: "twitter search", type: "api")
sandbase_inspect(name: "sandbase_twitter_web_search_timeline")
sandbase_run(name: "sandbase_twitter_web_search_timeline", arguments: {"keyword": "AI agents"})

Image generation

sandbase_discover(q: "flux", type: "multimodal")
sandbase_inspect(name: "sandbase_flux_schnell")
sandbase_run(name: "sandbase_flux_schnell", arguments: {"prompt": "A mountain lake at sunset"})

LLM inference

sandbase_inspect(name: "sandbase_openai_gpt_4o")
sandbase_run(name: "sandbase_openai_gpt_4o", arguments: {
  "messages": [{"role": "user", "content": "Explain quantum computing briefly"}]
})

Check recent costs

sandbase_runs(limit: 5)
→ [{ "model": "openai/gpt-4o", "cost": "0.000325", "status": "completed" }, ...]

Rules

1. Discover first — always verify a tool exists before running it.
2. Inspect before run — read the inputSchema. Never guess parameters.
3. Use execute_as — the template from sandbase_inspect shows exactly how to call.
4. Respect the user's stack — don't replace their existing tools.
5. Start small — use small limits on first calls for scraping/search tools.
6. Poll async runs — use sandbase_run_get for long-running operations.
7. Report costs — mention pricing when the user cares about budget.
8. One call per turn — wait for results before the next call.