How does Capman differ from MCP?

Capman generates capability manifests at build time by analyzing your TypeScript application's existing API surface. It requires no separate protocol server and produces a static, versioned manifest your agents can consume via any HTTP client. It is a lightweight MCP alternative focused on TypeScript applications.

How do I install Capman?

Install via npm: npm install capman. Then generate your manifest with: npx capman generate --from openapi.json (from an existing OpenAPI spec) or npx capman generate --ai (AI-assisted) or npx capman init (from scratch).

v0.6.2 — Concurrency Safety & Health API

The capability manifest your AI agent needs to stop guessing

Capman instruments your TypeScript application at build time, emitting a machine-readable capability manifest — a precise, versioned index of every API your app exposes. Your AI agent reads the manifest once and executes with surgical precision: no DOM traversal, no brittle selectors, no hallucinated click paths.

Start Building View on GitHub

The Problem

AI agents navigating your app by sight is a fragile architecture

DOM Dependency

Your agent parses Tailwind classes and infers intent from visual hierarchy. A class rename or layout refactor silently breaks the entire interaction chain with no warning..

Structural Brittleness

Every design iteration is a regression. Agents relying on CSS selectors and DOM traversal cannot distinguish a visual change from a functional one — both cause silent failures in production.

Hallucinated Click Paths

Without a capability map, LLMs invent navigation paths — clicking through five screens to reach a function that could have been a single direct API call with a typed response.

The Technical Edge

Build-time instrumentation.
Zero UI dependency.

Capman runs as part of your build pipeline. It analyzes your application's TypeScript surface — components, routes, API handlers — and emits a structured capability manifest: a precise, versioned index of everything your app can do.

Your AI agent reads the manifest once and executes with surgical precision. No DOM traversal. No brittle selectors. No guessing. Direct API execution against a schema it understands completely.

Without Capman


                agent → clicks Home → scrolls → finds button

                → navigates /events → parses DOM table

                → clicks row → reads modal → extracts value

                → breaks on next deploy

With Capman


                agent → reads manifest

                → calls getEventAvailability({ id, date })

                → receives typed response → done

Implementation

Zero to manifest in seconds

One package. No runtime overhead. Drop it into any TypeScript project and your AI agents immediately have a map.


            $
            npm install capman

Documentation

Everything you need to build

README.md

Overview, quick start, install steps, and core concepts. Start here.

Read overview

CODEBASE.md

Full technical reference — every option, every field, every internal decision.

Explore internals

CONTRIBUTING.md

How to contribute, dev setup, and PR guidelines. We welcome help.

Start contributing

FAQ

Common questions

What is a capability manifest?

A capability manifest is a machine-readable JSON file that describes everything your application can do — every API endpoint, its parameters, data access scope, and semantic description. AI agents read the manifest once and execute API calls directly, without navigating your UI. Capman generates and maintains this manifest automatically at build time from your TypeScript source.

How does Capman compare to MCP?

Capman is a lightweight MCP alternative focused on TypeScript applications. It generates capability manifests at build time by analyzing your existing API surface — no separate protocol server, no runtime process, no new infrastructure. The result is a static, versioned manifest your agents consume via any HTTP client. MCP requires a dedicated server; Capman requires a build step.

Does Capman add runtime overhead?

No. Capman runs entirely at build time. The generated manifest is a static JSON artifact. There are zero runtime dependencies added to your application. The CapmanEngine that ships with the package is consumed by your agent, not your app — your app's bundle is untouched.

Which LLMs does Capman support?

Any LLM that accepts a string and returns a string. Capman's LLM interface is model-agnostic — pass Anthropic Claude, OpenAI GPT-4, or any local model wrapper. The LLM reranker only fires for genuinely ambiguous queries; BM25 and fuzzy matching handle the majority of traffic without any LLM call.