SandBase CLI is an open-source AI command-line tool and local MCP (Model Context Protocol) server, with the goal of providing a unified external capability access layer for AI programming assistants and agents. It can connect 25 kinds of clients such as Claude Code, Cursor, Codex, ChatGPT, Hermes, and OpenClaw to 2000+ AI models and APIs, enabling AI assistants originally limited to the local environment to search the web, access social media, generate images, or obtain real-time data, without having to manually configure APIs one by one.
In terms of main capabilities, SandBase CLI works through an MCP bridge; after connecting, clients can discover, inspect, and run available SandBase models and APIs. It provides a sandboxed tool invocation flow: use sandbase_discover to find candidate endpoints by task or capability, use sandbase_inspect to view the current input schema, pricing, and execution requirements, and after confirming, invoke through sandbase_run. Data API calls are free, while model calls are billed by usage; inspect will display the current price, and base_price: 0 means free. For asynchronous tasks (such as image, video, and audio generation), a run_id will be returned, and sandbase_run_get can be used to obtain the result, avoiding duplicate task creation. In addition, sandbase_runs is provided to view recent run status and recorded costs, and sandbase_account to view the account balance.
In terms of usage, users can paste a set of instructions to an AI agent, letting it read the SandBase Skill and run connect on its own, then complete a one-time authorization login in the browser. They can also manually execute the installation command, and macOS or Linux users can also install through the official Homebrew formula. After authorization, every supported client can discover and run SandBase models and APIs through the local MCP bridge, and a new session needs to be started to load the sandbase_* tools. For agents that do not support MCP or have abnormal MCP sessions, the same capabilities can be used through shell transport, with credentials automatically provided by connect, or set through SANDBASE_API_KEY or sandbase keys add -k <key>.
In terms of security and manageability, SandBase CLI uses the OAuth device flow with PKCE, avoiding exposure of keys in URLs or command-line arguments; credentials are stored locally with 0600 file permissions; updates preserve existing JSONC comments and user-managed MCP entries; on failure or unregister, only SandBase-owned configuration is removed, achieving precise rollback; users can revoke authorization with one click at any time in the SandBase Dashboard.
Applicable scenarios include: supplementing AI programming assistants with external capabilities such as web search, social intelligence, image generation, video generation, and real-time data access; sharing unified model and API access across multiple AI clients; and agent workflows that require auditability and rollback. In terms of limitations, some features depend on a SandBase account and network connection; model calls may incur costs, and the price must be confirmed before running; OpenClaw currently does not support the Skill removal command; if api.sandbase.ai is unreachable, the bridge will retain SandBase tools and display actionable error information.
