Amplitude cut its MCP server from 96 tools to a handful after one connection ate 57% of a window

There's a new layer of tooling showing up in design and build work, and it is not the flashy "generate a whole screen" demo. It is smaller and more useful: tools that carry context for you so your team spends less time on upkeep and more time on calls that matter. Figma shipped an agent that lives on the canvas. Claude Code got better at remembering your rules. Amplitude rebuilt its MCP server and learned hard lessons about what agents can actually do. Let me catch you up.

The deep cut

• Context beats prompts. Figma's agent reads your attached library so designers stop retyping the same rules every session.
• Maintenance is design work, not cleanup. Uber publishes component docs in an afternoon that once took a team months.
• Cut tools down to what agents finish. Amplitude's 96 tools ate 57% of a 200K window before a single call.

The busywork you keep putting off

The Figma agent's real pitch is upkeep, not creation. It can audit a screen against your library, flag a "Save for later" button someone built from scratch instead of using Button/Secondary, and catch a hard-coded color where a token should be. It can reconcile naming when the same state is called hover, hovered, and mouse-over, and record what changed so the renames carry into code.

This is the drudge work your team avoids, which is exactly why systems drift. Figma's own framing puts it plainly: small inconsistencies pile up, erode trust, and push builders to work around the system instead of using it. Handing that off is worth more than any one generated mockup.

The payoff shows up in time, not magic

The proof is in named teams. At Uber, Staff Product Designer Ian Guisard built skills like /create-api and /create-color, and documentation that once took multiple people months now gets published by a single designer in an afternoon. Granola's Paavan Buddhdev pulls meeting notes onto the canvas and asks for 30 ways to phrase a button, calling it "a massive time saver."

The sharper lesson sits in Figma's Workflow Lab: the agent inspects, repairs, and distills feedback while the designer makes the design, product, and policy calls. When two components could both fit, the agent stops and annotates the frame instead of guessing. That line, what needs judgment versus what just needs doing, is the one to draw with your team.

Rules you store beat rules you repeat

The same idea runs through the code side. One Claude Code user admitted he typed the same instructions every session, use pnpm not npm, run the tests, then got annoyed when the model drifted. The fix was to store rules in three files: CLAUDE.md for context, settings.json for permissions the model can't talk around, and SKILL.md for workflows you run by name. Guidance is a suggestion. An enforced hook is a wall.

Figma's version of this is library guidelines: markdown a library owner uploads so the agent gets the rules automatically when the library is attached. Nobody retypes "only use red for Delete, never decoration." The context travels with the work.

Fewer tools, finished jobs

Amplitude's MCP rebuild is the warning label for anyone rushing to wire agents into their stack. By June their server loaded 96 tools, and one customer on a 200K window saw 57% of it eaten just by connecting, before a single call. More tools also meant more wrong picks. Even Opus 4.5 chose the right tool only 88% of the time.

The rebuild was about matching tools to how agents actually work. Cohorts went from nine tools to one. They stopped mirroring their API and built around outcomes, because agents are bad at chaining calls in the right order. Of 18,703 users who queried data, 20% rendered a chart and under 5% saved it, the chain broke. They watched recovery rate, whether an agent got to a good call within five minutes of a failed one, over raw error rate. Retention grew. A smaller catalog brought in more people.

Three questions for your team

• Where is our design system drifting right now, and could an agent audit for off-token values and duplicate components this week?
• What rules do we retype every session, and which belong in an attached library or an enforced config instead of a prompt?
• If we expose tools to an agent, are we measuring whether it finishes the job and recovers from failure, not just whether a call succeeds?