An AI / MCP API client

Tiger ships a built-in Model Context Protocol server. Point any MCP-compatible AI assistant at your collection and it can list, inspect and execute your requests without leaving its own interface - entirely on your machine, no account required.

What is MCP?

Model Context Protocol (MCP) is an open standard that lets AI assistants call external tools in a structured, auditable way. Instead of asking the model to guess how to invoke an action, you expose a typed set of tools it can call directly. The model sees input schemas and return types; it cannot do anything outside those boundaries.

Tiger exposes four MCP tools that cover reading collection structure and executing requests. There are no tools to write or delete requests, so the AI can explore and invoke but cannot modify your collection files.

Common use cases

The value is letting an assistant work with the requests your team already maintains in Git - with their real URLs, auth, variables and environments - instead of guessing or pasting curl into the chat. Typical uses:

• Agentic testing and debugging - "run the create-user request against staging and explain why it returns 401."
• Chained workflows - "call login, take the token, then run /me," combining run_request with Tiger's response captures.
• AI coding sessions - in Claude Code or Cursor, the model exercises your real endpoints through curated, environment-aware requests while it writes code.
• Onboarding and exploration - "list the requests and summarize what this API does" on an unfamiliar collection.
• Smoke checks and triage - have the assistant run the health-check requests and report anything that is not 2xx.

The four MCP tools

The server registers exactly these tools. Each has a typed input schema so the model can call them without guessing argument names or shapes.

How to connect Tiger to Claude Desktop

Tiger's MCP server is a compiled Node.js module (server.mjs) that runs as a local stdio process. The steps below use Claude Desktop as the example, but the same mcpServers block works in any MCP-compatible client (Cursor, etc.).

Step 1 - Build the MCP server

If you are running Tiger from source, build the server once:

npm run build:mcp

If you downloaded the packaged app, the compiled server.mjs is already included inside the app bundle at Tiger.app/Contents/Resources/out/mcp/server.mjs on macOS or resources/out/mcp/server.mjs on Windows.

Step 2 - Register Tiger in your AI client's config

Open Claude Desktop's config file. On macOS it lives at:

~/Library/Application Support/Claude/claude_desktop_config.json

Add the tiger block under mcpServers. Adjust the two paths to match your machine:

{
  "mcpServers": {
    "tiger": {
      "command": "node",
      "args": [
        "/Applications/Tiger.app/Contents/Resources/out/mcp/server.mjs",
        "/path/to/your/collection"
      ]
    }
  }
}

The second argument is the path to the Tiger collection directory you want the AI to access. The MCP server can only read files inside that directory; everything else on your machine is out of scope.

Step 3 - Restart the AI client

Restart Claude Desktop after saving the config. The four Tiger tools will appear in the model's available tool list on the next conversation. You can verify they loaded by asking the model "What Tiger tools do you have?" - it will enumerate list_requests, get_request, run_request and list_environments.

Step 4 - Ask your assistant to work with the API

The model can now list every request in your collection, read any request's full definition including headers, body schema and variable references, and fire a request against your target environment. Results (status, timing, response headers and a pretty-printed body) flow back into its context window so it can interpret the response, chain follow-up requests, or surface errors - all without you switching windows.

What the AI can and cannot see

The MCP server is designed around least privilege.

Secrets are redacted

Variables marked as secret in Tiger (shown with a masked value in the UI) are never sent to the MCP client. The AI sees the variable name, confirming the slot exists, but cannot read its value. This means the model can tell you "the request uses {{token}}" without receiving the actual token.

Read and run only

There are no MCP tools to write, rename, move, or delete requests. The AI can explore the structure of your collection and invoke requests, but it cannot modify any .tiger file on disk through the MCP interface.

Scope-limited by path

The collection path you pass in the config is the only directory the server can access. Other collections, other folders and other files on your machine are unreachable through the MCP tools.

Entirely local

The MCP server runs as a local process on your machine. There is no Tiger cloud service. Communication is only between your AI client and the local Node.js process. When run_request fires an HTTP call, that request travels directly from your machine to the target API - the same network path as pressing Send in the Tiger UI. No proxy or intermediary is added.

Why git-native storage makes this practical

Cloud-based API clients cannot safely expose a local MCP server because their collections live in a vendor cloud behind authentication. A local process that reads collection files directly only works when those files are actually on disk. Tiger's git-native, plain-text storage model is what makes the MCP server practical: the same .tiger files you commit, branch and review in pull requests are the files the AI assistant reads.

The table below reflects publicly documented features as of June 2026.

Example prompts to try

Once the tools are connected, these prompts work well as starting points:

• "List all the requests in my collection and group them by folder."
• "Show me the full definition of the Create Post request, including its body schema."
• "Run the Login request using the staging environment and tell me the response."
• "Run the Login request, capture the token from the response, then run the Get Profile request with that token."
• "Which requests use the {{userId}} variable?"
• "Find any requests that don't have an Authorization header set."

Related pages