vibekit

``` █ █ █ █▄▄ █▀▀ █▄▀ █ ▀█▀ ▀▄▀ █ █▄█ ██▄ █ █ █ █ ``` VibeKit is a toolkit for building Algorand agents: typed queries and actions, an agent loop with an HTTP handler, an x402 paywall, REST and MCP adapters, a keystore signer, and copy-paste React components — plus a CLI that sets up your coding agent and a terminal Explorer to watch it work. The [web agent](https://agent.getvibekit.ai) is built from these pieces and nothing else. ``` @initlabs/vibekit queries, actions, plugins, network, defineQuery, defineAction @initlabs/vibekit/actions the draft → approve → sign → confirm machine, wallet signing @initlabs/vibekit/agent createAgent, createAgentHandler @initlabs/vibekit/pay createPaywall (x402 → credit → a turn per call) @initlabs/vibekit/rest createRestHandler (POST /tools/<tool>) @initlabs/vibekit/mcp stdio and streamable HTTP vibekit add <component> companion, action, approval, tool-result, transaction, asset, account ``` Docs: [getvibekit.ai/docs](https://getvibekit.ai/docs) — start with [build an agent](https://getvibekit.ai/docs/tutorials/build-an-agent/). The reference implementation is [`apps/reference`](apps/reference): ~200 lines of app over the toolkit. The [web agent](https://agent.getvibekit.ai) is the product built the same way. Works with **Claude Code**, **Codex**, **Cursor**, **Copilot**, **Grok**, **opencode**, and **pi**. Alpha release — [feedback welcome](https://github.com/initlabsai/vibekit/issues). ## Requirements | For | You need | | --- | --- | | VibeKit CLI | Nothing — a self-contained binary | | VibeKit Explorer | Nothing — a self-contained binary | | Keystore | Node.js, and your OS keychain | | Starter templates | Node.js 24+ | | LocalNet | Docker Compose v2 | The keystore is what signs, so payments, assets, and deploys go through it. Its daemon runs under Node and is installed once with `npm`. On Linux it needs a Secret Service keychain; on Windows, the [Visual C++ Redistributable](https://learn.microsoft.com/en-us/cpp/windows/latest-supported-vc-redist) (`winget install Microsoft.VCRedist.2015+.x64`). The starter templates compile and test with `puya-ts`, `tsx`, and `vitest`. On Linux, LocalNet also needs you in the `docker` group (`sudo usermod -aG docker $USER`, then log out and back in). `vibekit doctor` reports what is missing without changing anything. ## Quick start macOS and Linux: ```bash curl -fsSL https://getvibekit.ai/alpha | sh ``` Windows: ```powershell irm https://getvibekit.ai/alpha.ps1 | iex ``` Both install the CLI and the Explorer sidecar. There is no stable release yet, so `/alpha` and `/alpha.ps1` are `/install` and `/install.ps1` with the prerelease channel pre-set. Scaffold a project and set up your agent: ```bash vibekit new my-app ``` Start a local chain and open the Explorer: ```bash vibekit localnet start vibekit explore ``` Then open your AI tool in the project and start building. ## Platform support | Platform | Status | | --------------------- | ------ | | Linux (x64) | Alpha | | macOS (Apple Silicon) | Alpha | | macOS (Intel) | Alpha | | Windows (x64) | Alpha | Development runs on Linux. The macOS and Windows binaries are built in CI and have been run by hand; they are less exercised than Linux. ## Why VibeKit AI coding assistants are bad at Algorand. They hallucinate APIs, reach for outdated patterns, and cannot actually deploy or test anything. `vibekit new` and `vibekit init` fix that. They install **skills** that teach your agent current Algorand patterns, and wire up **MCP tools** that let it touch the chain directly. Your agent can write a contract, deploy it to LocalNet, mint test assets, call methods, and check the result. Ask it what went wrong and it reads the chain to find out. Keys never reach the model. Signing goes through a local keystore daemon over a socket; the agent asks for a signature and never sees key material. Who reviews a write depends on the host. In the Explorer, every write stops at an approval screen showing the decoded group and its simulation. `vibekit mcp` defaults to execute mode and signs without prompting — set `SIGNING=compose` to get unsigned transaction groups back for a wallet to review instead. ## The Explorer `vibekit explore` opens a full-screen terminal Explorer: a chat transcript with a live results feed. Type an address, asset id, or transaction id and it resolves without a model call. Ask a question and the agent answers alongside trusted result cards. Contracts get a card each — creator, state, bare actions, live global state — and you can call any ABI method from the line, with reads simulated inline and writes routed through an approval modal that shows decoded arguments before you sign. ## Documentation Full documentation at **[getvibekit.ai](https://getvibekit.ai)** - [Your first project](https://getvibekit.ai/docs/tutorials/first-project) - [Explore with VibeKit](https://getvibekit.ai/docs/tutorials/explore-with-vibekit) - [Add VibeKit to a project](https://getvibekit.ai/docs/guides/add-to-an-existing-project) - [How VibeKit works](https://getvibekit.ai/docs/explanation/how-vibekit-works) ## CLI commands ```bash vibekit new [dir] # Scaffold a project, then set up agents vibekit init [dir] # Set up agents in an existing project vibekit explore # Open the Explorer TUI vibekit localnet <cmd> # Manage the local Algorand network vibekit doctor # Diagnose setup problems (--fix repairs them) vibekit keystore <cmd> # Signing accounts and the keystore daemon vibekit dispenser <cmd> # TestNet dispenser session vibekit tool <name> [json] # Call any VibeKit tool from the shell vibekit mcp # Run the MCP server over stdio ``` Templates for `vibekit new`: `contracts`, `fullstack`, `kitchensink`. ## Package The CLI is one host over a shared tool contract. The same tools ship as one TypeScript package, `@initlabs/vibekit`, published to npm under the `alpha` tag: ```bash bun add @initlabs/vibekit@alpha ``` `.` is the tool contract (also at `./core`); `./tools`, `./tools/views`, `./preset`, `./mcp`, `./mcp/stdio`, `./mcp/http`, `./agent`, `./agent/config`, `./signer-keystore`, and `./plugins/<name>` are the rest. `algosdk` and `zod` are the only required peers. Each subpath that wraps a third-party SDK declares it as an optional peer, so you install what you use: | Subpath | Install alongside | | --- | --- | | `.`, `./tools`, `./tools/views` | nothing more | | `./mcp`, `./mcp/stdio`, `./mcp/http` | `@modelcontextprotocol/server` | | `./agent` | `ai`, plus `@ai-sdk/anthropic`, `@ai-sdk/openai`, or `@ai-sdk/openai-compatible` | | `./signer-keystore` | `@algorandfoundation/keystore-core`, `@algorandfoundation/keystore-node`, `@tanstack/store` | | `./plugins/nfd` | `@txnlab/nfd-sdk` | | `./plugins/alpha-arcade` | `@alpha-arcade/sdk` | | `./preset` | all of the above | ## Development [AGENTS.md](./AGENTS.md) is the map of the repository: what runs where, how a tool call flows, the glossary, the rules, and the release procedure. ```bash bun install bun run build bun run typecheck bun run test bun run cli -- --help ``` - [Constitution](./docs/CONSTITUTION.md) — the bets the project rests on - [Contributing](./CONTRIBUTING.md) — conventions and review ## License MIT