01
Bring an objective
You (or your agent) hand the registry the real intent — a paragraph, not a keyword. The richer the brief, the better the menu.
Katailyst is the building-block library and orchestration layer your AI fleet can speak to from anywhere. Skills, knowledge, prompts, tools, recipes, rubrics, agents — curated, versioned, connected. Served once, consumed many times.
The cost of spinning up a new AI app or agent has collapsed. The cost of keeping ten of them coherent hasn't. Every workspace ends up with the same skills re-implemented in three places, the same knowledge stale-pasted in another, the same tool defined twice.
Katailyst is the answer to that drift. One registry, served via MCP, that any agent — Claude Code, Cursor, Codex, OpenAI Agents, Render-hosted Mastra workflows, your own Next.js app — can call to discover the right building block for the moment.
You bring an objective. The registry decomposes it, walks the graph, and hands you a kit. You decide what to use.
You (or your agent) hand the registry the real intent — a paragraph, not a keyword. The richer the brief, the better the menu.
Complex intents get split into 2–10 sub-questions. Each sub-question runs a vector search against the right region of the graph.
Sub-agents walk outward from each match, gathering related skills, recipes, knowledge, and rubrics. They return a kit, not a single answer.
Use 50%, 80%, or 100% of the kit. Mix it with your own context. The registry hands clubs to the player; it never swings.
Twenty-one entity types, grouped by what they answer. Capability says what an agent can do. Knowledge says what to know first. Voice + form says what the output looks like. Quality + measurement says how we'll know it worked. Operations holds it all together.
{
"mcpServers": {
"katailyst": {
"url": "https://www.katailyst.com/mcp",
"transport": "streamable-http"
}
}
}import { MCPClient } from '@modelcontextprotocol/sdk'
const mcp = new MCPClient({
url: 'https://www.katailyst.com/mcp',
auth: { bearer: process.env.KATAILYST_PAT }
})
const kit = await mcp.callTool('discover', {
intent: 'social post for nursing audience...'
})registry.capabilities returns the full system map — toolsets, prompts, resources, current org context.discover with a rich intent paragraph. Returns ranked entities + a discovery diagnostic so you can refine.We are the caddy, never the player.
The agents reaching for the registry have more situational context than the registry itself. They know who their user is, what time it is, what just failed, what the org's constraints are. Forcing them down a single pre-scripted route would waste that context.
So we don't. We curate the building blocks. We enrich the metadata. We tighten the ranking. When discovery returns the wrong thing, we fix the data — never wire a forced route. Agents pick. Agents compose. Agents own the swing.
That posture scales. Ten more apps spin up tomorrow. They all reach for the same registry. The registry gets better; every consumer gets better with it.
Open the dashboard, browse the registry, or wire up MCP and let your agent take the first turn.