It is code you own
Your agent is real TypeScript in a private repo you own, shown back as an editable graph. The manager is code, not an LLM.
Launch Agent Studio from your terminal.
Run Agent Studio locally and open its browser workspace to shape and inspect agents beside your code.
Agent Studio
npx @sapiom/agent-studio@latest
Starts locally and opens the Agent Studio workspace in your browser.
You build the agent. Sapiom runs everything it needs.
One integration, every capability
Search, scraping, browser automation, compute, data, storage, messaging, images and audio through one Sapiom key, without separate vendor keys.
Inspect every attempt
Input, output, directive, error, retry state, timing, and cost stay attached to each recorded step attempt.
One server to build with, one to call.
Both speak MCP and answer to the same account. What differs is where each one runs.
The authoring server
Scaffolds, validates the graph offline, runs it on stubs, and deploys when it is green.
$ mcp add sapiom-project -- npx -y @sapiom/mcp
The capability server
Every capability as a direct tool over HTTP, and the only way in for a hosted client.
$ claude mcp add sapiom --transport http https://api.sapiom.ai/v1/mcp
Two functions, and every capability hangs off ctx.
An agent is defineAgent({ name, entry, steps }). Each step calls capabilities on ctx.sapiom and returns the step that runs next.
The gate is in the graph, not in a comment.
The gate step approve? parks until digest.approved arrives, then resumes on the same state, in the same trace. Both arms end in a recorded state.
Sapiom Runtime
Planned sequenceresearch-brief · 4 steps · 1 waitStepSub-agentWaits
01collect-sourcesSearch
02read-sourcesExtract
TypeCapability callInputweb.scrapeDurationAt run timeEst. cost$0.010 / extraction
03Approve?Wait
04publish-digestSend
Every capability, available through one account.
Discover and call Sapiom capabilities from remote MCP or a typed agent step, with usage attributed to the agent and trace.
Web and search
Search the webRemote MCP
Search the web and return answers with sources or raw results.
sapiom_searchReference
Web scrapingRemote MCP
Scrape, crawl, map, or extract structured content from the web.
sapiom_scrapeReference
Deep researchRemote MCP
Research across sources and return a structured, cited result.
sapiom_deep_searchReference
Browser and compute
Extract contentTyped step
Extract structured fields from a page.
Crawl a siteRemote MCP
Follow a site and return its relevant pages.
Run codeRemote MCP
Execute code through the Sapiom compute surface.
sapiom_run
Create a sandboxRemote MCP
Start an isolated environment for agent work.
sapiom_sandbox_create
No capabilities match.
Your first agent, running in minutes.
Start in your coding agent, terminal, or Agent Studio. Sapiom handles the setup.
01
Start from a template
Clone an agent and run it locally.
02
Bring your own agent
Check, link, and deploy the TypeScript you have.
03
Start from intent
Describe the job and scaffold a typed agent.
Agent Studio is the desktop workspace: AI chat, Canvas, Steps, and TypeScript in one place.
npx @sapiom/agent-studio@latest Starts locally and opens the Agent Studio workspace in your browser.
mcp add sapiom-project -- npx -y @sapiom/mcp Added MCP server "sapiom-project". Run sapiom_authenticate once to sign in.
mcp add sapiom-project -- npx -y @sapiom/mcp Added MCP server "sapiom-project". Restart Codex to load its tools.
Everything you need to go deeper.
Read the docs
Quickstart, capabilities, Sapiom Runtime, schedules, and signals, kept current with what ships.
Open the docsSDK reference
Every argument of defineAgent and defineStep, every directive, and the full ctx.sapiom surface.
Read the referenceGitHub
The packages, the MCP server, and the template gallery. Read the source, file an issue, send a patch.
Star on GitHubShip your first agent today.
Build and validate locally for free. Production capability calls are metered by their published units.