SandBase

@sandbaseai/cli

Give your AI agent superpowers. One line. 2,000+ AI models and APIs. 100 free data API calls a day, then pay per call from one balance.

English | 中文 | 日本語 | 한국어 | Español | Français | Deutsch | Português

[GitHub release] [CI] [Official MCP Registry] [Indexed on TensorBlock MCP Index] [License] [GitHub stars]

[SandBase CLI connects AI clients to 2,000+ AI models and APIs]

Your AI coding assistant is smart, but it's trapped in a box. It can't search the web, check social media, generate images, or access real-time data — unless you wire up each API yourself.

SandBase changes that. One command connects your agent to 2,000+ AI models and APIs through the Model Context Protocol. No API keys to manage. No configuration headaches.

Building directly over HTTP instead of MCP? Start with the SandBase API quickstart or compare the unified LLM, image, and video API with specialized model gateways.

Setup

Paste this into your agent and let it do the rest:

set up https://www.sandbase.ai/SKILL.md

It reads the SandBase Skill, runs connect itself, and asks you to approve one browser sign-in. Prefer to drive it yourself:

npx -y @sandbaseai/cli@latest connect

Or install the official Homebrew formula on macOS or Linux:

brew install sandbaseai/tap/sandbaseai-clisandbase connect

See the Homebrew launch announcement and macOS/Linux CI evidence.

Approve the browser authorization once; each supported client can then discover, inspect, and run available SandBase models and APIs through the local MCP bridge. Start a new session afterwards so the sandbase_* tools load.

Every GitHub release publishes its own SHA-256 next to the tarball, so you can pin and verify an exact build instead of tracking latest:

V=$(npm view @sandbaseai/cli version)curl -fL -O "https://github.com/sandbaseai/cli/releases/download/v$V/sandbaseai-cli-$V.tgz"shasum -a 256 "sandbaseai-cli-$V.tgz"   # compare against the checksum on the release pagenpx -y "./sandbaseai-cli-$V.tgz" connect

Want to inspect compatibility before signing in or changing configuration? The client catalog is a read-only command — it writes nothing and needs no account:

npx -y @sandbaseai/cli@latest catalog --json

[Read-only SandBase CLI catalog audit showing 25 client targets] [SandBase CLI doctor failing closed for an empty isolated profile]

_{A no-write catalog audit and a bounded doctor check against an empty isolated profile.}

Homebrew users can run the same read-only check with sandbase catalog --json.

What It Costs

sandbase inspect <tool> always shows the current price before you run anything; data APIs covered by the daily allowance also show pricing.free_quota. The 100 free calls are shared across eligible data APIs and reset at 00:00 UTC; after that, calls are billed per call. sandbase balance --json shows today's remaining free calls as free_api_quota. Prices can change, so trust the live response over any example in this file.

Visit the official SandBase CLI project page for a compact overview. For a security-focused walkthrough of client detection, MCP tools, and rollback, read the 25-client SandBase CLI setup guide. AI agents and automated installers can use the concise llms-install.md instructions.

SandBase Open-Source Stack

• SandBase Harness — self-hosted agent runtime with persistent sessions, sandbox isolation, approvals, audit, and replay.
• SandBase Agent Skills — 88 installable Skills for research, social intelligence, marketing, and business workflows.
• SandBase for Codex — official Codex plugin packaging the SandBase MCP bridge and guided Agent Skill; its required HOL security scan passes on main.

Discover the authenticated remote bridge in the official MCP Registry, read its Registry community showcase, or find it through Chinese Independent Developer, Awesome MCP, Glama, TensorBlock MCP Index, MCPRepository, VaultPlane, AIMCP, MCP Server Hub, MCP Server Spot, or the Awesome AI API Proxy. External maintainers have merged SandBase into Awesome Codex CLI, Awesome Agent Skills, and Awesome MCP Server. The Hermes integration is independently listed in the Hermes Atlas ecosystem catalog and as beta in Awesome Hermes Agent. Install its verified sandbase Agent Skill from skills.sh, the sandbaseai/cli/sandbase listing on SkillsCat, the sandbaseai/cli/sandbase listing on skills.re, the 93/100 verified SandBase listing on vSkill, the security-audited SandBase listing on Skillstore, or the versioned sandbaseai/sandbase snapshot on Agent Skill Hub.

A Verifiable Agent Workflow

After connecting, ask your agent to follow the same inspectable sequence for search, data APIs, language models, image generation, video generation, and other catalog capabilities:

1. Use sandbase_discover to find candidates by task or capability.
2. Use sandbase_inspect to read the current input schema, pricing, and execution requirements.
3. Confirm the selected endpoint and parameters before calling sandbase_run. Data APIs include 100 free calls per organization per UTC day, then are billed per call like models; confirm the cost shown by sandbase_inspect before paid runs.
4. For asynchronous work, use sandbase_run_get with the returned run_id instead of creating a duplicate run.
5. Use sandbase_runs to review recent run status and recorded cost.
6. Use sandbase_account to check the current account balance.

For example, start with a non-billable discovery request:

Find image models suitable for a square product illustration. Compare the required inputs and current pricing for the best two candidates. Do not run either model yet.

Discovery and inspection never cost anything, so an agent can always survey the catalog before it spends.

After reviewing the live response, explicitly approve the endpoint and inputs you want to run. Catalog entries, schemas, pricing, latency, and availability can change, so use the tool response from the current session rather than a static example.

See the complete Claude Code and Codex walkthrough or the Chinese workflow guide. For a live catalog example, read the GitHub connector discover-and-inspect dogfooding note.

What's Inside

Supported Clients (25 Targets)

Works with every major AI coding tool:

Preview the complete compatibility catalog without signing in or changing any files:

npx -y @sandbaseai/cli@latest catalog --json

# Connect all detected clients at oncenpx -y @sandbaseai/cli@latest connect
# Or target one specific clientnpx -y @sandbaseai/cli@latest connect --client cursor

OpenClaw quickstart

Target OpenClaw directly, then ask the CLI to read back the managed MCP server and SandBase Skill before you rely on them:

npx -y @sandbaseai/cli@latest connect --client openclawnpx -y @sandbaseai/cli@latest doctor --client openclaw

After doctor reports the credential, MCP bridge, and Skill as configured, restart OpenClaw and make a natural-language request that needs SandBase. If OpenClaw already has a non-SandBase MCP server named sandbase, the installer leaves it untouched instead of overwriting it. To remove the managed MCP entry and local credential later:

npx -y @sandbaseai/cli@latest unregister --client openclaw

OpenClaw currently exposes no supported Skill removal command, so unregister reports that limitation and leaves the verified SandBase Skill installed.

Hermes quickstart

SandBase can install both its managed Hermes Skill and local MCP entry in one authorization transaction:

npx -y @sandbaseai/cli@latest connect --client hermesnpx -y @sandbaseai/cli@latest doctor --client hermes

doctor verifies the owned local credential, MCP entry, and Skill. Restart Hermes to confirm native discovery in a fresh client session, then use the SandBase Skill or MCP tools. A pre-existing, non-SandBase MCP entry named sandbase is left untouched. To remove only the managed Hermes configuration, Skill, and local credential while retaining the shared bridge for other clients:

npx -y @sandbaseai/cli@latest unregister --client hermes

The integration is independently listed by Hermes Atlas and Awesome Hermes Agent.

How It Works

┌─────────────────┐         ┌───────────────┐         ┌────────────────────┐│   Your Agent    │  MCP    │  Local Bridge │  HTTPS  │    SandBase API     ││  (Cursor, etc.) │────────▶│  (on-demand)  │────────▶│ 2,000+ models/APIs ││                 │◀────────│               │◀────────│                    │└─────────────────┘         └───────────────┘         └────────────────────┘

1. Run connect → CLI opens browser auth → you approve
2. API key saved locally (file permissions 0600)
3. MCP bridge configured for your client
4. Agent calls tools on-demand. No daemon. No background process.

Commands

CLI Commands

sandbase connect [--client <name>]    # Authorize + configuresandbase setup                        # After upgrading: refresh the local bridge + Skill, test connectivitysandbase doctor [--client <name>]     # Local checks plus one real round trip to SandBasesandbase unregister [--client <name>] # Remove configurationsandbase catalog --json               # List all supported clients

Agent Commands (shell transport, same capabilities as the MCP tools)

Agents without MCP, or with a broken MCP session, use the same catalog from a shell. Credentials come from connect automatically (or SANDBASE_API_KEY / sandbase keys add -k <key>).

sandbase discover -q "twitter posts"                       # search; -t api|llm|multimodal|embedding|servicesandbase inspect sandbase_twitter_web_search_timeline       # fields, price, ready-to-run commandsandbase run sandbase_twitter_web_search_timeline -i '{"keyword":"AI agents"}'sandbase run <video_tool> -i '{"prompt":"..."}' -w -o result.json   # wait with backoff, save output to disksandbase runs get -r <run_id> -w -o result.json            # resume a long run instead of re-running itsandbase upload ./photo.png                                # local image/audio/video -> public URL for tools that take a URLsandbase runs list | sandbase balance

-j prints JSON; errors are {"error":{"code","message","hint"}} with the next step in hint. Exit codes: 0 success or still running, 1 failure (including a failed run), 2 usage.

Sync tools (LLMs, most data APIs) return their result directly; async tools (image, video, audio jobs) return a run id — wait with -w, or fetch later with sandbase runs get -r <run_id>. If api.sandbase.ai is unreachable, the bridge keeps the SandBase tools visible with an actionable error instead of dropping the server.

MCP Tools (available to your agent after connecting)

Recommended workflow:

discover → inspect → run

1. sandbase_discover(q: "twitter search")          → find current candidates2. sandbase_inspect(name: discovered_name)          → get schema and pricing3. sandbase_run(name: discovered_name, input: {...}) → run after confirmation

Security

• Zero secrets in URLs or CLI args — OAuth device flow with PKCE
• Restricted file permissions — Credentials stored with 0600
• Ownership-aware updates — Existing JSONC comments and user-managed MCP entries are preserved
• Exact rollback — On failure or unregister, only configuration owned by SandBase is removed
• Revoke anytime — One click in the SandBase Dashboard

Get Started

npx -y @sandbaseai/cli@latest connect

Then ask your agent to do something it couldn't before.

Create an account → — every organization gets 100 free data API calls a day.

If SandBase saves you setup time, star the repository — it helps more agent users discover the project.

Links

• SandBase Platform
• Documentation
• CLI setup guide
• CLI deep dive: English · 中文
• Six-tool practical workflow: English · 中文
• npm Package
• Official Codex Plugin
• Dashboard

Community

• Ask a question or share what you built
• Report a reproducible bug
• Propose an integration or improvement
• Contribute to SandBase CLI

Community-curated listings:

• Official Registry community showcase
• VaultPlane MCP directory
• AIMCP directory
• Awesome MCP
• Awesome AI API Proxy
• Awesome Gemini CLI
• Hermes Atlas
• Awesome Hermes Agent
• Awesome AI Tools
• Awesome MCP Server
• Awesome Agent Cortex
• Awesome AI Agent Tools
• Awesome Codex CLI

License

Apache-2.0