You've invested years building your API. Adding AI shouldn't mean starting over. Most of your API's power is locked behind UIs only power users navigate. Give every customer a way in.
Three interfaces from one API spec.
Connect your OpenAPI spec. Pick your interfaces. Ship.
Chat assistant
Your customers ask questions. Your API answers them.
The agent maps natural language to the right endpoints automatically. No docs for your users to read, no tickets for your team to answer.
CLI tools
Power users get CLI tools they can script and chain.
Every API capability becomes a command. Pipe it, chain it, build workflows. Works with local AI agents and MCP clients too.
Platform integration
Your product works inside ChatGPT, Claude, Cursor, and more.
It becomes a native tool in the AI platforms your customers already use. No custom integration work on your end.
From OpenAPI spec to working agent.
Connect your API
Point it at your spec
$ pikku agent init --spec https://api.acme.dev/openapi.json
✓ Parsed 42 endpoints
✓ Generated function types
✓ Created agent project
Next: pikku deploy
Pick your interfaces
Chat, CLI, MCP — any combo
// Your API endpoints become agent tools automatically
const agent = pikkuAIAgent({
name: 'acme-assistant',
instructions: 'Help users manage their account',
tools: [getDeployments, createIncident, getCloudSpend, ...],
})
// Expose as chat, CLI, and MCP
wireChannel({ channel: 'support', agent })
wireCLI({ program: 'acme', agent })
wireMCP({ agent })
Ship it
Deploy to Fabric
One AI call per workflow. Zero tokens after that.
The AI figures out your workflow once and turns it into a deterministic execution plan. Every run after that is native code. No model calls, no token costs, no latency variance.
LLM on every request
~$14,400/yr
8-step workflow, 50x/day
Pikku (AI designs once)
~$1.20/yr
Same workflow, native execution
Your API's rules still apply.
Permissions carry over
Your API's auth layer stays in the loop. The agent can only do what the user could already do.
AI can't go rogue
It designs the workflow. It doesn't execute it. Your existing permissions enforce every request.
Full audit trail
Every conversation, workflow, and execution is logged.
Deterministic execution
Approved workflows produce the same result every time. No temperature, no variance.
You built the API.
Let your customers use all of it.
Upload your OpenAPI spec. Get a chat assistant, CLI tools, and MCP server — deployed on Pikku Fabric.