Skip to content
webvise

Keel

Built in-house

Keel makes a SaaS product usable by AI agents. It reads the product's API, interface and docs, turns every operation into a typed tool on a hosted MCP server and puts one endpoint in front of them. Every call runs inside scopes and limits, a write that trips a guardrail waits for a person's approval, and each call leaves a trace that can be replayed.

Industry
AI Agents / Agent Infrastructure
Location
Berlin, Germany
Duration
1 week
3 Open standards it builds on
Keel: Keel: one guarded endpoint that lets AI agents use a SaaS product
4 Kinds of control per tool

The Challenge

SaaS products are built for people. A button, a form or a help article works for someone who can see and click. An AI agent needs the same action as a typed tool it can plan with, and an API alone does not give it that. The agent still cannot tell which calls are safe or whose permission it acts with. One API key gives it all rights or none, and a call without a record cannot be reviewed or repeated. Teams close the gap with hand-written glue code for each integration, and that code falls behind whenever the product changes.

The Solution

Keel works in a loop of five steps. First it maps the product. It reads the API, the interface and the docs and finds every operation, such as creating an invoice. Then it generates a typed tool for each operation, with a name, a description, an input schema and the scopes it needs, and hosts all tools on an MCP server. When the product changes, Keel maps and generates again, so the tools match the product. Next it guards each tool. An agent holds only the scopes it was granted, and Keel blocks any call outside them. Rate and spend limits apply to each tool. A write that trips a guardrail, such as an invoice above a set amount, is held until a person approves or denies it. Keel then publishes one endpoint and docs written for agents. Agents sign in once and call the tools over MCP, and a company's own services call the same tools over REST. Last, Keel records every call with its inputs, scope and outcome. A team can replay a call and see who made it and who approved it. Connectors bring in the products a team already runs, such as Stripe, Salesforce, Intercom, Linear, Slack and GitHub.

Project details

Tech Stack

Next.js, React, TypeScript, Tailwind CSS, Turborepo, Vercel