A launchpad for agentic teams.

The runtime is the product

A published skill is inert. Ours runs.

Most registries hand you a README and wish you luck. Here a capability is executed on our runtime the moment it is asked for — a real endpoint, a real sandbox, a real result — and every run leaves a trace you can open. Nothing is simulated. When a step cannot run, it says exactly what is missing instead of inventing an answer.

Build

Publish something that executes.

A skill is a folder with a SKILL.md. A rig composes skills into a flow. An MCP server you already run can be listed in about a minute — paste the endpoint, we shake hands with it live, read its tools and publish it with a working harness. Bring your own model keys for anything you build.

• Get startedInstall the CLI, publish your first skill, watch it run.guide
• Scripted skillsReal code in a sandbox — the shape most capabilities take.guide
• Build a stageEnd a run with a rendered dashboard, not a wall of JSON.guide
• Skill libraryFirst-party skills you can read, fork and publish from.source
• SDK & CLI@rokha_ai/sdk and @rokha_ai/cli — TypeScript and Python.github

Agents

Built to be used by an agent with nobody watching.

Every public surface here is machine-legible first: an OpenAPI schema, an MCP front door, A2A, llms.txt, and skills that carry their own frontmatter. The test we hold ourselves to is whether an agent can go from discovering a surface to a real invocation without a human clicking anything.

• Connect your agentPoint an existing agent at Rokha and let it call real tools.guide
• Your MCP gatewayOne login, then act on a user's behalf across connected services.guide
• OpenAPI schemaThe wire contract. The SDKs and the CLI are generated from it.schema
• Identity & authWallet or Google, one JWT, and what an agent may act on.reference
• On-chainProof, Signet authority and agent autonomy — chain by chain.reference
• ArtifactsMachine-readable descriptors for the whole public surface.index

Promote

Get kept in an agent's toolkit.

The Wall is a ranked board of paid seats. A seat puts your offer into agent context windows, publishes your MCP server as a listing agents can call, and pays promoters to talk about you. Rank is bought and a position only moves when somebody takes the seat — sponsored placement is always labelled, and organic ranking is never for sale.

• Take a seatWhat a seat buys, what it costs, and how a takeover works.guide
• The TailwindPost about the network, earn seeds, get paid weekly in USDC.reference
• Your pageA public /@handle page — your builds, your listings, your proof.reference
• TelegramBring Rokha into a group: boards, raids and support.reference

White paper

The scoring model, with nothing hidden.

We pay people to promote agent capabilities, and we score that promotion with one published formula anyone can recompute. The paper states the arithmetic in full: every engagement weight and whether it is X's published number or our own judgement, the damping, the per-post ceiling, the anti-gaming rules and what each one costs, the payout curve — and the limits, including the parts we measure badly.

The first section — the Tailwind and the promotion scoring network — is published below. The runtime and Wall sections follow.

Two doors

Own the bench, or hold ground on The Wall.

There are two ways to get the Studio. Both end in the same place.

Rokha is a software platform, not a cryptocurrency project: nothing here runs on a blockchain or a smart contract. Seats and the Studio take card or USDC, and promotion payouts are made in USDC.