Sapiom SDK

⚠️Beta Status: Currently in v0.x (beta). API may change before v1.0.0. Production-ready and actively maintained.

TypeScript SDK for building, running, and operating AI agents on Sapiom. Author agents as typed step graphs, call Sapiom paid tools (sandboxes, git repos, coding models, search, file storage, …) directly from your code, and ship them to the Sapiom engine from the CLI, your coding agent's MCP, or the Sapiom Studio desktop app.

📦 Packages

This is a monorepo of focused packages. Install only what you need.

Build & run agents

Package	Version	Description
@sapiom/agent		The authoring contract: defineAgent, defineStep, directives (goto/terminate), and types
@sapiom/tools		Typed client for Sapiom capabilities — the same tools your agents call, callable from your code
@sapiom/cli		Command line: scaffold, validate, deploy, and schedule agents
@sapiom/mcp		Local developer MCP server (sapiom-dev) — build & operate agents from your coding agent

Agent Studio

Agent Studio runs your coding agent (Claude Code or Codex) in a Sapiom-configured environment: MCP pre-wired, agent projects tracked, one-click deploy/run, and a live canvas for previews.

Package	Version	Description
@sapiom/agent-studio		Agent Studio launcher — npx @sapiom/agent-studio@latest
@sapiom/harness		The Agent Studio implementation — a CLI-launched local web app
@sapiom/harness-desktop		Sapiom Studio desktop app (Electron) — ships as signed installers, not to npm

Runtime internals

Lower-level packages that power the stack above. Most users never import these directly, but they're published for advanced/host integrations.

Package	Version	Description
@sapiom/agent-core		Pure, stateless functions for scaffolding, validating, and operating agents — shared by the CLI and MCP
@sapiom/agent-runtime		Host-agnostic graph-walker runtime — one runtime, two hosts (server engine + local runner)

🚀 Quick Start

New to Sapiom? The fastest path is the CLI or the developer MCP — both scaffold a working agent for you. See the examples folder for complete, runnable projects.

Scaffold an agent with the CLI

npx @sapiom/cli agents init my-app # scaffold a projectcd my-app
npx @sapiom/cli agents check # validate locally (bundle, manifest, graph)
npx @sapiom/cli agents deploy # build and ship

Build with your coding agent in Agent Studio

One command checks your environment, signs you in, and opens Agent Studio with your coding agent running in an embedded terminal:

npx @sapiom/agent-studio@latest [dir]

Prefer a native app? Sapiom Studio is the one-click desktop host for the same experience — download installers (macOS, Windows, Linux) from GitHub Releases.

Author an agent

An agent is a typed graph of steps. Each step does work and returns a directive telling the runtime where to go next.

import{defineAgent,defineStep,goto,terminate}from"@sapiom/agent";conststart=defineStep({name: "start",next: ["finish"],asyncrun(input,ctx){returngoto("finish",{greeting: `hello ${input.name}`});},});constfinish=defineStep({name: "finish",next: [],terminal: true,asyncrun(input){returnterminate({done: true, ...input});},});exportconsthello=defineAgent({name: "hello",entry: "start",steps: { start, finish },});

Call Sapiom capabilities from your code

@sapiom/tools exposes the same capabilities your agents call as tools, typed and authenticated to your tenant.

import{createClient}from"@sapiom/tools";constsapiom=createClient({apiKey: process.env.SAPIOM_API_KEY});// Create a repo, have a coding model build into it, then publish.constrepo=awaitsapiom.repositories.create("landing-page");construn=awaitsapiom.models.coding.run({task: "Build a one-page marketing site in index.html.",gitRepository: repo,});if(run.result?.success){const{ sha }=awaitrepo.pushFromSandbox(run.sandbox,{message: "build: landing",});console.log("published",sha);}

Build agents from your coding agent (MCP)

Add the local developer MCP so your coding agent can scaffold, test, deploy, and inspect Sapiom agents. In Claude Code:

claude mcp add sapiom-dev -- npx -y @sapiom/mcp

@sapiom/mcp is the local developer surface (sapiom_dev_*). It is distinct from the remote Sapiom capability MCP that services paid tool calls — see docs/mcp-servers.md for which to use when.

📚 Documentation

• Examples — runnable example projects
• The two Sapiom MCP servers — local dev vs. remote capabilities
• @sapiom/agent — authoring contract
• @sapiom/tools — capability client
• @sapiom/cli — command line
• @sapiom/mcp — developer MCP
• Agent Studio — local app for building with your coding agent

🏗️ Package Architecture

@sapiom/agent Authoring contract (defineAgent, directives, types)
↑
├── @sapiom/agent-runtime Host-agnostic graph-walker runtime
└── @sapiom/agent-core Scaffold / validate / operate (pure functions)
↑
├── @sapiom/cli Command line
└── @sapiom/mcp Local developer MCP (sapiom-dev)
@sapiom/tools Typed capability client (sandboxes, repos, models, …)
@sapiom/harness Agent Studio (local web app; launched via @sapiom/agent-studio)
↑
└── @sapiom/harness-desktop Sapiom Studio desktop app (Electron host, ships as installers)

🛠️ Development

This is a pnpm workspace monorepo.

# Install dependencies
pnpm install
# Build all packages
pnpm build
# Run tests
pnpm test# Lint and format
pnpm lint
pnpm format

Package Scripts

# Build / test a specific package
pnpm --filter @sapiom/agent build
pnpm --filter @sapiom/tools test# Watch mode
pnpm --filter @sapiom/agent dev

Publishing

We use Changesets for version management:

pnpm changeset # create a changeset
pnpm version-packages # apply version bumps
pnpm release # build and publish to npm

🤝 Contributing

Contributions welcome! Please read our Contributing Guide first — it explains which changes can go straight to a pull request (focused bug fixes, documentation corrections, single-template additions) and which need a maintainer-agreed issue before you invest in them (new features, public API changes, new dependencies, cross-package work).

In short:

1. Fork the repository and branch from the latest main
2. Keep the change focused on one problem, with tests for changed behavior
3. Run the root checks (pnpm build, pnpm typecheck, pnpm lint, pnpm test)
4. Add a changeset when a published package's behavior or API changes
5. Open a pull request and complete every applicable section of the template

AI-assisted contributions are welcome but must be disclosed, reviewed, and validated by the contributor. For suspected security vulnerabilities, follow the Security Policy instead of opening a public issue.

📄 License

MIT © Sapiom

🔗 Links

• Website
• Documentation
• NPM Organization
• GitHub Issues