How to Use Thunder Client's MCP Server: The Beginner's Guide
CodexFuse Team · 2026-11-03T18:20:30.396Z · 12 min read min read
What Is Thunder Client's MCP Server?
Learning how to use Thunder Client MCP means teaching your AI coding assistant to drive your API client by chatting: Thunder Client's MCP server lets your AI assistant create, update, and run Thunder Client requests and collections from plain-language prompts, according to Thunder Client's own documentation.
Thunder Client itself is a fast, privacy-first API client for VS Code and JetBrains IDEs whose requests and collections are stored locally and shared through your own Git repository, never a third-party cloud. Its docs call it the most popular API client on the VS Code Marketplace, with 7M+ installs. One clarification up front, because three products share the Thunder name: this guide covers Thunder Client's MCP server only. Thunder Compute is a GPU cloud with its own MCP server, and Thunders is an AI-native testing platform — both different products, documented separately. If you want to sanity-check a live endpoint while following along, CodexFuse's free REST API tester runs in the browser with no signup.
What Do You Need Before Installing?
You need three things: Node.js 22 or later, the Thunder Client extension installed and up to date, and the Thunder Client CLI version 2.1.0 or later installed globally. The CLI is the bridge the MCP server talks to, so the version floor matters.
npm i -g @thunderclient/cli@latest tc --version
Run tc --version and confirm it reports 2.1.0 or newer before going further. If tc is not found after installing, the npm global bin directory is probably not on your PATH — the fixes are in the troubleshooting section below.
How Do You Install the Thunder Client MCP Server?
Installation is a five-step arc that takes a few minutes: install the CLI, register the server with your AI tool, restart, verify the connection, and run a setup check. The vendor's install docs currently show the exact steps for Claude Code; the overview page separately lists Codex, GitHub Copilot, Cursor, Antigravity, Kiro, and Windsurf as supported, with the "works with any MCP-compatible tool" promise, but only Claude Code's exact command was verified at write time.
- Install the CLI globally. Run
npm i -g @thunderclient/cli@latest, then confirm withtc --version(2.1.0 or later). - Register the server with Claude Code. Run this in your terminal:
claude mcp add thunderclient -- npx -y thunderclient-mcp
This fetches the open-source, MIT-licensed server package (thunderclient-mcp on npm) on demand. - Restart Claude Code completely. The new server only loads on a fresh start.
- Verify the connection. Run
/mcpin a Claude Code session;thunderclientshould be listed as connected. New to the tool? The Claude Code beginner guide covers the client side from scratch. - Test with a setup check. Ask your assistant: "Check that Thunder Client is set up for my project." It reports the CLI version, the data path it is using, and your collections and environments — if anything is missing, the troubleshooting section below tells you where to look.
What Can the AI Assistant Actually Do With Your API Project?
The assistant gets thirteen tools plus two reference guides, grouped by job. Understanding the groups matters more than memorizing names, because it tells you what kind of sentence to write.
| Job | Tools | What each does |
|---|---|---|
| Inspect your workspace (read-only; never changes data) | debugInfo, listCollections, listEnvironments, getDetails | Check setup and return CLI version, data path, collections, and environments; list collections with folders and requests; list environment names and which is active; return full details of a collection, folder, or request by ID. |
| Create and update | createCollection, createFolder, createRequest, updateRequest, updateCollectionSettings | Create a collection (optionally with a folder); add a folder to a collection (one level of subfolders); create a request from a curl command, with optional auth, tests, scripts, and docs; update a request's name, URL, method, headers, parameters, body, auth, tests, scripts, or docs; configure a collection or folder's base URL, shared auth, default headers, tests, scripts, environment, and run options — or rename it. |
| Run requests and collections (sends real HTTP) | runRequest, runCollection | Run one request and get its status, response body, and test results; run a collection or folder for a per-request report with status, response time, and test results. Both use the active environment unless you name another in your prompt; runs support iterations, delay, concurrency, stop on failure, and a specific environment. |
| Delete | deleteRequest | Permanently deletes a request from a collection or the Activity tab. It cannot be undone. |
| Reference guides (the assistant reads on demand) | tc://docs, tc://env-guide | Tool overview, URL and base URL conventions, the token-capture pattern, and a scripting API reference; guidance on using environment variables and handling secrets safely. |
Two gotchas the docs flag: when updateRequest sets headers, parameters, or tests, it replaces the existing list instead of merging, so the assistant must include every item that should remain. And runCollection redacts query parameter values in its reports. If you like to audit what an AI agent is allowed to do before connecting it, the MCP Census guide shows a vetting workflow for MCP servers generally.
What Does a Real AI-Driven API Test Look Like?
A realistic arc moves from setup check to generated tests to a staging smoke run, all in sentences. The vendor's example prompts page demonstrates each of these workflows; adapt the collection names, URLs, and file paths to your own project.
- Setup check. "Check that Thunder Client is set up for this project, then summarize my collections and environments, and tell me which environment is active."
- Generate a collection from source code. "Analyze the Express routes in src/routes/ and create a 'Shop API' collection with one folder per resource. Set the base URL to http://localhost:3000, use path-only request URLs, and include a realistic sample body for each POST and PUT request." (There is also an import-from-OpenAPI variant: "Read openapi.yaml and create a 'Payments API' collection that covers every endpoint, grouped into folders by tag.")
- Configure auth. "Configure the 'Billing API' collection to use OAuth 2.0 client credentials with the token URL https://auth.example.com/oauth/token,
{{clientId}}and{{clientSecret}}, and the scope 'billing:read billing:write'." - Add baseline tests. "Add tests to every request in the 'Orders' folder: the expected success status code, response time under 1000 ms, and a JSON body that contains an 'id' field."
- Chain an end-to-end flow. "Create an order, save the returned id to
{{orderId}}, fetch the order by{{orderId}}, update its status to 'shipped', then delete it." The token-capture pattern behind this is documented in thetc://docsresource. - Run a smoke test against staging. "Run the 'Shop API' collection using the 'Staging' environment and give me a summary of the results, grouped by folder, with any failed tests listed first."
- Debug a failure. "GET /orders/{id} in the 'Shop API' collection returns 401. Compare its auth and headers with a request in the same collection that succeeds, explain the difference, and fix it."
Four prompting habits the docs recommend: describe the outcome, not the tool ("Check my setup" works better than naming a tool); use exact collection and folder names as they appear in Thunder Client; ask for a plan first on large changes; and review deletions before approving them, because deleted requests cannot be recovered. Secrets stay safe in this flow: reference values as {{varName}}, since the assistant sees variable names but not their values — values themselves are set in Thunder Client's Environments view.
What Can't the MCP Server Do? The Real Limitations
The MCP server is deliberately a subset of the full client. Anything on this list must be done in the Thunder Client UI instead; knowing the boundary up front saves you from arguing with the assistant.
| Task | Where it has to happen |
|---|---|
| Create, edit, or activate environments — and set any environment variable values | Thunder Client UI |
| AWS and NTLM auth (credentials cannot be set through MCP) | Thunder Client UI |
| Delete collections and folders (only individual requests can be deleted via MCP) | Thunder Client UI |
| Move requests between collections or folders, or save Activity-tab requests into a collection | Thunder Client UI |
| Binary request bodies (file uploads) | Thunder Client UI |
| SSE, WebSocket, and gRPC requests (cannot be created or run through MCP) | Thunder Client UI |
Is It Safe to Let an AI Assistant Run Your API Tests?
The privacy model is the reassuring part of the design. The docs state there is no analytics or usage tracking, and the assistant never sees environment variable values or saved credentials — only their names are shared, and the values are masked. Responses longer than 50,000 characters are truncated before they reach the assistant, and query parameter values are redacted in runCollection reports.
The one honest caveat: response bodies are shared with the assistant, so the docs advise avoiding requests that return data you do not want to share. The assistant also cannot read or modify your environment files directly — it can only use the masked variable names.
Thunder Client MCP Server Not Working? Fixes for the 6 Common Errors
The install page's troubleshooting table covers the failure modes beginners actually hit. Each error maps to exactly one fix.
| Error | Fix |
|---|---|
tc not found | Install the CLI globally with npm install -g @thunderclient/cli, make sure the npm global bin directory is on your PATH, and restart your AI tool afterward. |
| No JSON output from CLI | Your CLI is older than version 2.1.0. Update it with npm install -g @thunderclient/cli@latest. |
projectDir must be an absolute path | Give the full path to your project, such as /Users/john/my-project, not a relative path like ./my-project. |
| No collections found | Ask the assistant to check your setup. Confirm the reported data path matches where Thunder Client saves your data; if you use workspace settings, see the VS Code Settings doc. |
{{varName}} values don't resolve | Create an environment in Thunder Client → Environments, add the variable, and set it as active. |
| Server not listed in your AI tool | Check the configuration file for JSON syntax errors, then fully restart the AI tool. |
Sources
The facts above were verified against these sources, all fetched on November 3, 2026: the Thunder Client MCP install page (docs.thunderclient.com/mcp/install — requirements, the claude mcp add thunderclient -- npx -y thunderclient-mcp command, the /mcp verify step, and all six troubleshooting entries); the MCP tools reference (docs.thunderclient.com/mcp/tools — all thirteen tool descriptions, the two resources, the limitations list, and the privacy model); the example-prompts page (docs.thunderclient.com/mcp/usage — every prompt workflow and the four prompting tips); the MCP overview page (docs.thunderclient.com/mcp — the "create, update, and run Thunder Client requests and collections from plain-language prompts" description and the supported-tools list); the docs home page (docs.thunderclient.com — the "fast, privacy-first API client for VS Code and JetBrains IDEs" description and the 7M+ install figure); and the VS Code Marketplace changelog for the Thunder Client extension (v3.1.0 "Improved MCP Server v2" dated 2026-10-27, the "Now supported on the Free Plan" note, the v2.37.4 original MCP server, and the github.com/thunderclient/thunderclient-mcp MIT-licensed repo link). All four docs pages carry the stamp "Last updated on November 3, 2026". Thunder Compute (thundercompute.com) and Thunders (thunders.ai) were checked only to confirm they are different products.
Frequently Asked Questions
Is Thunder Client's MCP server free?
The server itself is open source and MIT-licensed, and the extension changelog states the improved MCP server is "Now supported on the Free Plan" as of v3.1.0 (October 27, 2026). The docs give no subscription tiers or prices, so check the vendor's current plan page before relying on any paid feature.
Does the MCP server work with Cursor, Windsurf, or GitHub Copilot?
The overview page lists Claude Code, Codex, GitHub Copilot, Cursor, Antigravity, and Kiro, plus any MCP-compatible tool, and the changelog adds Windsurf. At write time only Claude Code's exact install command was verified in the docs, so confirm the other tools' setup steps in the vendor's current install page rather than guessing them.
Is Thunder Client the same as Thunder Compute or Thunders?
No. Thunder Client is the VS Code and JetBrains API client this guide covers. Thunder Compute is a GPU cloud platform with its own MCP server for managing instances, and Thunders is an AI-native testing platform with its own MCP server. Same "Thunder" name, three different products.
Can the AI assistant see my API keys and secrets?
No. The assistant sees environment variable names but never their values or saved credentials — the values are masked, and secrets are set in the Thunder Client Environments view, which the server cannot touch. One boundary to respect: response bodies are shared with the assistant, so don't run requests that return data you don't want to share.
Can the assistant delete my collections by accident?
It cannot delete collections or folders at all — the only deletion tool, deleteRequest, removes individual requests permanently. Even so, the docs recommend reviewing deletions before approving them, since a deleted request can't be recovered.
