SandBase CLI: One MCP Bridge for 25 AI Client Targets

AI clients increasingly support the Model Context Protocol, but each client still has its own configuration file, command, and lifecycle. Connecting the same tool surface to Codex, Claude Code, Cursor, Gemini CLI, or Windsurf can turn into a collection of one-off setup instructions.

SandBase CLI is an open-source TypeScript CLI and local MCP bridge that makes that setup repeatable. It detects supported clients, uses an OAuth device flow with PKCE, stores the resulting credential with restricted local permissions, and adds only the configuration owned by SandBase.

The current catalog covers 25 client targets and the MCP bridge exposes six focused tools for discovering, inspecting, and running items from a catalog of 2,000+ AI models and APIs.

Connect a detected client

Run the normal setup command:

npx -y @sandbaseai/cli connect

The CLI detects compatible clients and guides the setup for the clients it can configure safely. You can also select one target explicitly:

npx -y @sandbaseai/cli connect --client cursor

Before signing in or changing a local configuration file, inspect the complete compatibility catalog:

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

That command is useful in CI, support scripts, and security reviews because it returns machine-readable capability and verification metadata without writing to the machine.

Supported client families

The catalog includes auto-configured, guided, and manual integration paths for clients such as:

• Codex
• Claude Code and Claude Desktop
• Cursor and Cursor CLI
• Gemini CLI
• Windsurf
• OpenCode, Qwen Code, and Kimi CLI
• Kiro IDE and Kiro CLI
• Warp, Amp, Hermes, and OpenClaw
• ChatGPT and Cowork workflows that require workspace-level review

The distinction between automatic and guided setup is deliberate. If a client does not expose a stable configuration interface, the CLI prepares an owned artifact or prints the exact MCP launcher instead of pretending it completed a change it cannot verify.

Six MCP tools instead of a giant static catalog

After connection, the agent receives a compact workflow:

Tool	Purpose
sandbase_discover	Search the model and API catalog
sandbase_inspect	Read the input schema, pricing, and request template
sandbase_run	Execute a selected model or API
sandbase_run_get	Poll an asynchronous run such as video generation
sandbase_runs	Inspect recent calls and cost information
sandbase_account	Check account balance without starting a paid run

The intended sequence is:

discover -> inspect -> run

That sequence matters. Agents should inspect a tool’s current schema and price before executing it rather than guessing parameters from a name.

Local ownership and rollback

Agent configuration is security-sensitive, so the CLI uses ownership-aware updates:

• credentials are not passed through command-line arguments or URLs;
• stored credentials use restricted file permissions;
• existing JSONC comments and unrelated MCP entries are preserved;
• failed setup rolls back SandBase-owned changes;
• unregister removes only configuration owned by SandBase;
• doctor reads the installed state and reports what still requires a real client confirmation.

Run a health check at any time:

npx -y @sandbaseai/cli doctor

Inspect the project

The repository is licensed under Apache-2.0 and includes its release workflow, tests, client capability registry, issue templates, security policy, and the installable sandbase Agent Skill.

• Source code
• Current GitHub release
• Installable Agent Skill
• Report a setup problem

If the CLI saves you from maintaining another set of client-specific MCP instructions, star SandBase CLI on GitHub. That signal helps other agent builders find and evaluate the project.