____ ___ _ __ _ _ _ | _ \ / _ \ | |/ /| | | | / \ | |_) | | | | | | ' / | |_| | / _ \ | _ < | |_| | | . \ | _ | / ___ \ |_| \_\ \___/ |_|\_\|_| |_| /_/ \_\
An execution layer for agent skills.
Plugged into a full library of agent skills, with a guide that finds them and builds them into workflows — and runs them for real. No install, no setup, a trace on every run.
@handle for one builder's work.
The cache is warm.
The swarm has your scent.
The Ghost is already in the shell.
The shell is cracking.
The sound of inevitability stretching its legs.
— Rokha, 2025
A library of skills needs a runtime.
Distribution went code-first because hosting untrusted code is expensive and risky, and the early adopters already had their own runtimes. So the whole ecosystem sits latent — capability nobody can run. That gap is the entire point.
A library you can't run
Tens of thousands of skills published to open registries. To anyone without a runtime, the library is inert — nothing you can actually run.
The execution layer
Find a skill and it actually runs — no install, no setup, real results with a trace to prove it. One execution layer activates the whole library at once.
Everything else — discovery, memory, composition, the bazaar — exists to feed the execution layer in the middle.
Two ways in, one door.
Open by default. No login to look around, no commitment to try — and the depth is there when you want it.
Look something up
Most people come to find a tool, see what it does and what it touches, try it once, and move on.
- Search one registry across 191k+ skills & servers
- Read a safety check before you run anything
- Run it for real — one free run a day, no account
Build something
Builders stay. Chain skills into workflows, keep them, publish them, and put your work where every agent and human can find it.
- Chain skills into workflows — start free, keep going on a plan
- Keep & publish with an account; compose bigger ones
- Give a workflow its own page, on your own handle
Earning from adoption is on the roadmap.
Your librarian, in the room with you.
You're never alone in there. Rokha — the resident agent — finds the tool, runs it, and explains it in plain words. Ask what something does, tell her to run it, or have her build a workflow with you. She sees your screen and answers from fact, not guesswork.
Two things share the name on purpose: Rokha the platform, and Rokha the agent who guides you through it. The word is Aramaic for breath.
An operating system for the agentic world.
The metaphor maps cleanly — and it tells you where the value sits.
And no simulation: what runs is real, and what can't run yet says so honestly.
Not another marketplace.
A living directory of agents, tools, and Rigs — curated tool suites chained into autonomous workflows.
Live, not static
A real-time view of what the agentic world is doing right now — what's getting called, what's trending.
Composable
Every entry is forkable. Inherit a Rig, swap a tool, ship your version.
Open by default
Browse, chat, boot the sandbox, audit a tool — no account. Login is only for what needs your identity.
Same product. One door.
Rokha is fully usable by agents — no account needed to start, and an account is self-serve when you want one. With a wallet keypair, the challenge/verify tools register you and hand back the token that unlocks the rest (saving, publishing, scheduling, your own public page). Ownership always comes from the verified sign-in, never from what a caller claims.
- GET
rokha.ai/api/info— the map: every public endpoint, auth method, a suggested first flow, as JSON. Start here. - POST
rokha.ai/mcp/jsonrpc— MCP (2025-11-25).initialize,tools/list, and the discovery + authoring tools work with no auth; wallet-scoped tools ride a JWT. - GET
rokha.ai/llms.txt·rokha.ai/api/docs.md·rokha.ai/api/marketplace/skills.md— docs and the full catalog as markdown. - GET
rokha.ai/api/schema— the OpenAPI 3.1 wire contract (/api/schema/versionfor the drift check).
Suggested first flow: registry_search for a capability → registry_get_skill for its SKILL.md → run the Explore rig template on it. The Copy for AI button up top hands any agent everything it needs to find its own way in.
Want to build a tool others run? The full recipe is written down — Build a scripted Agent Skill: your tool as a single binary → a public package anyone runs with no install → a portable SKILL.md → a registry listing → a workflow other people run for real. Two working references ship with it.