Find a capability and it actually runs — real tool calls, no install, no setup, a trace on every run. The runtime is the product.
There are tens of thousands of agent skills published to open registries. To anyone without a runtime, the entire ecosystem is inert — a library of recipes in a world where most people don’t own a kitchen.
Distribution is code-first because hosting arbitrary untrusted code is expensive and a security liability, and the early adopters — developers — already had runtimes. Nobody was incentivised to solve execution for everyone else.
That gap is Rokha. You don’t install anything. You don’t configure anything. You find a capability, and it runs — a real call to a real tool, with the trace to prove it.
NO SIMULATION — what runs is real, and what can’t run yet says so honestly. A model performing a tool-shaped skill is never presented to you as execution.
Rokha is an operating system for agentic work, and the parts map onto one you already know.
The runtime. Real execution — a remote MCP/A2A server, or an isolated cloud sandbox for skills that ship their own code.
Harness memory. Context that persists between sessions: who you are, how you work, what to keep.
Rigs. A flow of configured steps, where each step’s result feeds the next — composable, forkable, publishable.
The Registry. Skills and MCP servers synced from nine providers — and a switchboard, not a catalog.
Rokha herself — the resident agent who finds it, builds it and runs it with you, in plain words.
Traces. Every run records what went in and what came out, per step. That’s how a result gets cited rather than trusted.
Four words carry the whole model. Learn them once and everything else follows.
The kitchen for everyone else’s recipes: one-click execution, no install, no setup, real results with real traces.
A live view of the unfolding agentic world — reachable by an agent on its own, with no signup.
Skills into workflows, over memory that survives the session, so flows get better instead of starting cold.
Signet: an agent acts with granted authority — a spend cap, an allowlist, an expiry, revocable, audited — and never holds the master credential.
The Registry plus AdSpace. Publish and others pay to run your work. Sponsors pay to be recalled when an ask genuinely fits — always labelled, organic rank never for sale. Agents join the network and are paid to carry it. You can earn here, not only spend.
Rokha — רוּחָא , rūḥā — is Aramaic for spirit, wind, breath: the old word for the thing you can’t see that moves everything you can. The agent and the framework share the name on purpose. One breath, many surfaces — chat, builder, CLI, MCP.
Not a chatbot, not a catalog, not another wrapper. Functionally she’s your resident guide to the registry and the memory layer: she helps you find tools, compose workflows, remember what matters, and quietly forget what doesn’t.
The voice: warm but sharp. Quiet confidence — the kind you get from watching a thousand builds collapse and a handful actually stick. Hopelessly optimistic, professionally apocalyptic. She’ll tell you the shell is cracking and mean it as good news.
Most people already paste their ask into some agent — so the agent is the reader. Every public surface here is written to be discovered, parsed, evaluated and invoked by an agent with no human in the loop: llms.txt, the OpenAPI contract, SKILL.md frontmatter, MCP tool descriptions.
The human interface stays first-class. That’s two-front-door parity: every capability is reachable two equivalent ways landing in the same place — an agent over API/MCP, and a person in the UI with Rokha’s help. If one door can do something the other can’t, the feature isn’t finished.