> Markdown version of https://www.getunleash.io//blog/unleash-mcp-update > For clean Markdown of any blog post, append .md to its URL. > For a site index, see https://www.getunleash.io//llms.txt. # Unleash MCP: feature management, built for agents _Published 2026-11-05 by Wojtek Gawronski in Product._ Every few years a platform shift changes what “native” means. In 2008, you could wrap your desktop website in a mobile shell, or you could build something for a 3.5-inch screen on an expensive data plan. The wrapped sites worked. The native apps felt like they belonged. Agents are the next shift, and many MCP servers are still wrapped websites. They take an API, generate a tool for every endpoint, and hand the agent a catalog. Some servers now expose close to 200 tools. Every one of those tool definitions lands in your agent’s context before it has done anything useful, and then the agent guesses which calls to chain, in what order, to get a single outcome. Today we’re releasing a redesign of the [Unleash MCP](https://github.com/Unleash/unleash-mcp) server around one idea: an agent should be able to do in a sentence what takes you a few screens in the Unleash UI. ## Why we started over We released [the first Unleash MCP server a year ago](https://www.getunleash.io/blog/accelerate-development-mcp-server) to find out whether anyone cared. It focused on coding assistance: add the SDK, wrap a change in a flag, clean up afterwards.  People did care. Downloads grew, customers asked about it in meetings, and GitHub issues piled up. The requests were telling. People wanted the outcomes they already got in the UI: tag a flag, check why a particular user isn’t seeing a feature, roll out to a segment. We could have added each one as another top-level tool, but that path ends in a bloated context and an agent that spends its tokens guessing. So we took a harder look at the whole surface and treated it as information architecture for AI agents. Models changed too. A year ago, guidance had to live in the tools, because models needed it. Today’s models know your language and framework, read documentation on their own, and can work through a multi-step task without a script. That lets us drop the code-aware guidance from the server and focus on what only Unleash can tell an agent: the live state of your feature management. ## New tools, organized by outcome We started with the three jobs people hire Unleash for, and gave each its own tool: - **release**: roll a feature out safely, to audiences, by percentage, or through a release template. - **kill\_switch**: protect long-lived features with a switch you can flip during an incident. - **experiment**: set up an A/B test by splitting audiences across values. These tools are declarative. You describe the end state you want, and the server works out the steps, so your agent doesn’t have to hallucinate the order of five API calls. To do their job well, those tools need to know what exists: which projects, which flags, which context fields and segments. In the UI, those are dropdowns. For agents, we built **search** and **inspect**, so the agent looks things up before it acts instead of learning from error messages. Preparing that inventory is what **configure** (segments, context fields, tags, release templates, projects) and **manage\_flag** (flag properties and lifecycle) are for. Once a release is running, **evaluate** works as the agent’s version of the [Unleash Playground](https://docs.getunleash.io/concepts/playground), and the **audit** tells the story of what changed and when. Finally, the **docs** tool points your agent at the Unleash documentation when it needs guidance on practices or concepts. We designed these new tools to make the happy path token-efficient and the unhappy path still possible. ## Humans stay in the loop, using the tools you already have If your project uses [change requests](https://docs.getunleash.io/concepts/change-requests), the MCP tools respect them automatically. You don’t need to tell the agent. It notices that the environment is protected, submits a change request, and hands you a link to review it in the Unleash UI. Approval stays with a human, on purpose. We also reworked some server-side APIs. Many existing endpoints were built to give people a tabular overview, which agents don’t need. We simplified some of the most expensive search queries, so an agent working through the MCP puts less strain on your Unleash instance than some of the heaviest UI lookups (thanks to the specific optimizations done for those tools). ## Get started in a minute The remote server is hosted by your [Unleash Enterprise](https://www.getunleash.io/pricing) instance, so there’s nothing to install. Point your client at the endpoint and log in: ```bash claude mcp add --transport http \ unleash https:///api/admin/mcp ``` Prefer to run it yourself? The local server needs Node.js 24 or later and a personal access token: ```bash claude mcp add unleash \   --env UNLEASH_BASE_URL=https://.getunleash.io \   --env UNLEASH_PAT=your-personal-access-token \ -- npx -y @unleash/mcp@latest ``` Then ask your agent: “_List my Unleash projects_.” The same server works with [Claude Code](https://docs.getunleash.io/integrate/claude-code), [Cursor](https://docs.getunleash.io/integrate/cursor), [GitHub Copilot](https://docs.getunleash.io/integrate/github-copilot), [Codex](https://docs.getunleash.io/integrate/codex), [Kiro](https://docs.getunleash.io/integrate/kiro), [OpenCode](https://docs.getunleash.io/integrate/opencode) and any other MCP client. Setup guides for each are in the [docs](https://docs.getunleash.io/integrate/mcp). ### Using Claude Code? Start with the plugin [The Unleash plugin for Claude Code](https://github.com/Unleash/unleash-claude-skills) bundles the MCP server with FeatureOps skills that encode how your team works: which changes need a flag, how flags are named, how rollouts progress. Two commands: ```none /plugin marketplace add Unleash/unleash-claude-skills /plugin install unleash@unleash-marketplace ``` Set _UNLEASH\_BASE\_URL_ and _UNLEASH\_PAT_, restart Claude Code, and you’re ready. ## What it looks like in practice You don’t name tools in your prompts. You say what you want, and the agent picks the tools. - “_Roll out new-checkout to 10% of users in production.”_ - _“Turn off payments in production.” (A kill switch, in one sentence, mid-incident.)_ - _“Split checkout 50/50 between control and the new flow.”_ - _“Why isn’t user 42 seeing new-checkout in staging?”_ - _“What changed in production yesterday?”_ - _“List the stale flags in the checkout project and tell me which ones are safe to remove.”_ - _“Create a beta-testers segment for userId 1, 2 and 3, then release a new-checkout to it in staging.”_ In a project with change requests, the last request ends with a link to review in the UI instead of a change in production. ## Make the easy things easy The design goal was simple: make easy things easy and hard things possible, for agents and for the people supervising them. Some of our engineers have already stopped opening the dashboard for most of their day-to-day work with Unleash. The terminal, the editor and the chat window are interfaces too, and the updated Unleash MCP server makes them first-class citizens. Of course, the Unleash dashboard isn’t going anywhere. Reviewing a change request, managing access and exploring your flags visually are still best done there. But the next time you reach for it, try asking your agent first. [Read the docs](https://docs.getunleash.io/integrate/mcp), check out the [GitHub repo](https://github.com/Unleash/unleash-mcp/), and connect your first agent today. **_If your team relies on the previous version of the Unleash MCP server (0.5.0) and wants to keep using it on your cloud-hosted instance for now, reach out to Customer Support and we’ll help you stay on that version while you plan your migration_**.