A platform for building things on Gonka inference.
Gonkablocks is the layer between your idea and a production deployment that runs on the Gonka decentralized GPU network. Pick a tool from the library, fork it, or push your own. Inference, hosting, public URL, metering, and sharing are all wired by default.
Anything that fits in a Docker container.
A block is a Docker image plus a small manifest.yaml that declares its inputs, outputs, and runtime. That's the whole contract — there are no other restrictions.
- Any language. Python, Node, Go, Rust, C++, Bash — if it has a Dockerfile, it's a block.
- Any framework. FastAPI, Express, llama.cpp, pandas, ffmpeg, Playwright, Selenium, headless Chrome — bundle whatever you need into the image.
- Any system dependencies. CUDA libraries, native compilers, shared objects, custom binaries — install them in the Dockerfile the same way you would on a server.
- Network is allowed by default. Block code can call any external API or scrape the web — the platform only controls how the block is invoked, not what it does internally.
Inputs are passed in as JSON via /in/inputs.json; the block writes results to /out/outputs.json and (optionally) any extra files. Stdout/stderr are streamed live to whoever is watching the run.
Six ways to run that container.
The same Docker image can be invoked under any of the following runtime profiles. Pick the one that matches the shape of your tool — the platform handles the lifecycle.
- Tools (jobs). One-shot. Read inputs, do work, write outputs, exit. Perfect for translate, summarize, transcribe, agent runs.
- Workflows. Visual graphs of blocks wired input-to-output. The platform schedules each node and streams events.
- Sessions (chats). Long-running interactive container with WebSocket I/O. Auto-shuts on idle.
- Services (HTTP APIs). Container exposes a port; gonkablocks gives you a stable public URL with auth and observability baked in.
- Cron. Same job container, fired on a cron expression you set in the manifest.
- Workers. Long-running daemons — queue consumers, watchers, schedulers.
Gonka inference is built into every container.
Every block process starts with OPENAI_API_KEY and OPENAI_BASE_URL already set. Use the OpenAI SDK as you normally would — calls go through the platform's inference proxy to the Gonka network. No keys to manage, no separate billing account, no rate-limit dance.
Live on Gonka today: Qwen3 235B and Kimi K2.6 — Moonshot's flagship agentic coding model with 262k-token context and native swarm orchestration up to 300 sub-agents. Image, video, embeddings and fine-tuning will land the same way — picking a different model is a one-line edit in the manifest.
Right now: free for everyone.
Runs, inference and compute are all sponsored by Gonka Labs while we polish the platform and grow the community. Every run is metered internally so we can see usage shape, but nothing is charged. When the time comes to switch on real pricing, it will use Gonka's on-chain token economics — the same numbers exposed on rpc.gonka.gg. No surprises.
CLI, IaC manifest, MCP server.
Install the CLI globally and you're one command away from a public block:
npm install -g @gonkalabs/blocks-cli gonkablocks init # scaffold a starter block gonkablocks deploy # build, push, register, expose
Multi-block apps are described in a gonkablocks.yaml IaC manifest. Forking a public block is one click in the UI or one CLI command.
There's also an MCP server — @gonkalabs/blocks-mcp — that lets Cursor, Claude or any coding agent fork blocks, deploy new ones and watch live runs on your behalf.
The library is built by the community.
Public blocks ship their source — anyone can read, fork, modify, and republish. That's how the catalogue grows: one person solves a problem once, and the rest of the community gets a clean starting point. Powering all of it is the same Gonka inference network, so a fork inherits the free runtime automatically.