Amitte in five minutes — what the product actually does
A guided tour of every surface, what each one is for, and the shortest path to seeing the value yourself.
The one-paragraph version
Amitte is an open registry where you discover, sign, version, and fetch three things AI agents need: skills (playbooks), agents (model-backed callables), and MCP servers (tool surfaces). You can use the registry as a consumer (pull and run published entities) or a publisher (ship your own). It's accessible through a portal (this site), a CLI, a TypeScript SDK, an HTTP API, and an MCP server an agent can connect to directly.
That's the elevator pitch. Now the tour.
The portal — what each route does
| Route | What it's for |
|---|---|
/search | The catalog. Filter by entity type, trust tier, capability. |
/s/<id> | A skill / agent / MCP server detail page. |
/compose | Describe a goal in English, get a publish-ready bundle. |
/publish | Ship a single entity. |
/publish/bundle | Ship skill + agent + MCP server together as one repo. |
/me | Your published entities. Edit, deprecate, revoke. |
/blog | This. Conceptual posts + tutorials. |
/docs | API reference + integration guides. |
/policy | Governance, takedown policy, content rules. |
Every detail page has the same tab order: Overview / Manifest / Tools or Endpoints / Versions / Evals / Dependencies / Graph. The Graph tab draws the relationships visually — handy when you're auditing a bundle or trying to understand what an agent depends on.
The CLI — @amitte-ai/cli
Three commands cover 90% of what people do at the terminal:
# Search the registry
amitte search "release notes"
# Pull a skill into your local project
amitte pull amitte/repo-conventions-checker
# Connect a remote MCP server as a local stdio shim
amitte connect amitte/registry-tools
The connect command is the one most teams discover late and love. It turns a remote MCP server into a stdio process your client (Claude Desktop, Cursor) treats like any local server.
The SDK — @amitte-ai/sdk-ts
Same surface as the HTTP API, typed, with first-class entity-type methods:
import { Skills } from '@amitte-ai/sdk-ts';
const client = new Skills({ registryUrl: 'https://registry.amitte.com' });
const results = await client.search({
query: 'release notes',
entity_type: 'agent',
trust_floor: 'verified',
});
const skill = await client.fetchSkill('amitte/repo-conventions-checker');
Drop it into a Node.js / TypeScript agent and the registry becomes a typed dependency.
The MCP server — amitte/registry-tools
The registry exposes itself as an MCP server. A connected agent gets three tools:
search_skills(q)— full-text + facet search.fetch_skill(id)— manifest + body markdown for one entity.compose_intent(goal)— ask the planner to draft a bundle for a goal sentence.
Connect once, and the model can browse the registry without you ever copying a curl command.
The shortest path to "I get it"
- Open
/search. Pick any skill that looks interesting. - On the detail page, click Download → markdown body. Paste it into Claude Desktop or your terminal.
- Watch the agent run a real, signed, versioned procedure.
That's the whole product, end to end. Everything else — the manifest schema, the trust tiers, the bundle wizard, the BYOK compose flow — is infrastructure to make step 1-3 boring and reliable.
Where to go next
- New consumer? Read First five minutes as a consumer next.
- New publisher? Read First five minutes as a publisher instead.
- Curious about why this product exists? Read Why Amitte exists.