Foundry Studio

Foundry Workflows

Build it your way. We'll keep it running.

Build an agentic workflow on your own machine, as a Claude Code project or any script. Add one small file, studio.yaml, and deploy it. Foundry runs it on a schedule, on a webhook or by hand, and asks a person before anything that matters.

curl -fsSL https://app.foundrystudio.dev/install.sh | sh

Then run studio login.

Animated demo in a terminal, in a project folder called receipts. studio workflows init finds a Claude Code project with a /file-receipts command and proposes a studio.yaml: run /file-receipts in Claude Code every day at 07:00 Europe/London and on a webhook, connect Gmail and Drive, and ask first before sending email. It writes the file. studio workflows dev runs it on the same machine with the same runner as the cloud: it reads 3 new emails in finance@example.com, files 2 receipts in Drive and drafts a reply to accounts@example.com, then waits for approval to send it. Once it is approved in Studio, the email is sent and the run finishes in 52 seconds. studio workflows deploy reports that the connections are ready and that one secret, XERO_TOKEN, is still needed. studio workflows secrets set XERO_TOKEN saves it, encrypted. A second studio workflows deploy prints: Live as v1. Next run 07:00 Europe/London.

One workflow, start to finish

From the inbox to Drive, every morning.

The receipts workflow from the demo. Your agent does the work, your code does the matching, and a person approves anything that goes to a supplier.

Diagram: the receipts workflow, start to finish The receipts workflow, from start to finish. It starts every day at 07:00 London time, or when its webhook is called. Inside its own container, which can reach only the hosts you list, your agent, a Claude Code project, reads the finance inbox through Gmail (an AI step), matches receipts with code (a code step) and files them to Drive (an AI step). When it wants to email a supplier, the flow branches: a person approves on their phone or the web (a person step), then the email is sent (an AI step). Either way, the run ends logged in Studio, with every step shown in a thread. Its own container · only the hosts you list Your agent · a Claude Code project TRIGGEREvery day 07:00Europe/London TRIGGERWebhookor run it by hand AIRead finance inboxthrough Gmail CODEMatch receiptsyour script AIFile to Drivereceipts saved Email a supplier? PERSONA person approveson phone or web AIEmail sentafter approval STUDIORun logged in Studioevery step in a thread Diagram: the receipts workflow, start to finish The receipts workflow, from start to finish. It starts every day at 07:00 London time, or when its webhook is called. Inside its own container, which can reach only the hosts you list, your agent, a Claude Code project, reads the finance inbox through Gmail (an AI step), matches receipts with code (a code step) and files them to Drive (an AI step). When it wants to email a supplier, the flow branches: a person approves on their phone or the web (a person step), then the email is sent (an AI step). Either way, the run ends logged in Studio, with every step shown in a thread. Its own container only the hosts you list Your agent: a Claude Code project TRIGGEREvery day 07:00Europe/London TRIGGERWebhookor run it by hand AIRead finance inboxthrough Gmail CODEMatch receiptsyour script AIFile to Drivereceipts saved Email a supplier? PERSONA person approveson phone or web AIEmail sentafter approval STUDIORun logged in Studioevery step in a thread
  • AI step
  • Code step
  • Person step
  • The run's own container

Today all three kinds work inside one run of your agent. Workflows built from separate agent, code and person steps, with agentic steps you can switch off, are on the way.

How it works

Build. Test. Deploy. It runs.

You build with the tools you already use. Foundry does the part that never ends: running it every day, and telling you when something needs you.

  1. 01

    Build

    Work in Claude Code as you do now: CLAUDE.md, skills, subagents, slash commands, MCP servers and its permission settings. Or write any script, such as one on the Claude Agent SDK. Add studio.yaml.

  2. 02

    Test

    studio workflows dev runs it on your machine with the same runner the cloud uses, against your real connections.

  3. 03

    Deploy

    studio workflows deploy, or ask Claude Code to deploy it: our Claude Code plugin knows how. It tells you when it runs next and which connections or secrets it still needs.

  4. 04

    It runs

    On a schedule, on a webhook or by hand. Each run's steps show in a Studio thread. If a run fails, you get a push.

One small file

studio.yaml says how it runs.

Your code stays as it is. This file sits beside it and says what to run, when, and what it may touch.

studio.yaml
version: 1
name: receipts
description: File receipts from the finance inbox to Drive
entry:
  claude: "/file-receipts"
triggers:
  - cron: "0 7 * * *"
    tz: Europe/London
  - webhook: {}
connections: [google.gmail, google.drive]
secrets: [XERO_TOKEN]
egress: [api.xero.com]
approvals:
  ask: ["mcp__studio__gmail_send"]
limits: { timeout: 10m }
state: state/
entry
What to run. Here, a slash command in a Claude Code project. It can be any script instead.
triggers
Every day at 07:00 London time, and whenever the webhook is called. You can always run it by hand too.
connections
Gmail and Drive, connected through Studio. No keys in your repo.
secrets
Names only. You set the values with studio workflows secrets set; they are stored encrypted.
egress
The only hosts a run can reach, besides the model and Studio.
approvals.ask
Tools that wait for a person. This run can draft an email, but not send it on its own.
limits
A run stops after 10 minutes.
state
A folder kept from one run to the next.

Run it for good

It keeps running, or it tells you why not.

Durable runs

Built on our durable workflow engine. Runs survive crashes and deploys, retry cleanly and never silently stop.

Every step in a thread

Each run's steps show in a Studio thread, so you can see what it did. If a run fails, a push tells you.

Versions and rollback

Every deploy is a new version. studio workflows rollback takes you back to an earlier one.

Logs, pause and resume

studio workflows logs shows what happened in each run. Pause a workflow and resume it when you are ready.

In charge and contained

People decide. Runs stay in their box.

People stay in charge

  • Anything listed under approvals.ask waits for a person, such as sending an email or pushing code
  • Approve from the web or from your phone
  • Each run's steps show in a Studio thread, so you can see what it did
  • A failed run sends you a push

Safe by default

  • Each run gets its own container
  • A run can reach only the hosts you list, plus the model and Studio
  • Your model key never enters the container
  • Every run has a time limit
  • Connections without handing over keys: Gmail, Drive, GitHub and more, through Studio, with personal or company accounts
  • Secrets are set with studio workflows secrets set, stored encrypted and never written to files

Open by design

Yours to keep, and to take elsewhere.

Your repo

The workflow lives in your own repository, built with your own tools.

A plain YAML file

studio.yaml is short and readable. Nothing proprietary goes into your code.

Your model key

Bring your own model key. Anthropic is the default; use any model you like. Model usage is billed to you by your provider.

MCP and Claude Code

Your Claude Code project comes as it is: CLAUDE.md, skills, subagents, slash commands and MCP servers.

On the way Not available yet

  • Import and export in open standards: Open Agent Specification and CNCF Serverless Workflow 1.0
  • Multi-step workflows where each step is an agent, code, a tool call or a person. Agentic steps can be switched off where you need certainty, each with a fallback
  • Workflow diagrams (AWML)
  • Graphs brought in from the Rivet visual editor
  • A visual editor in Studio

Pricing

Start free.

Free

Everything you need for your first workflows.

  • 3 workflows
  • 300 run-minutes a month
  • Runs up to 10 minutes
  • Schedules as often as every 15 minutes
  • Webhooks and manual runs
  • Approvals, logs and rollback
  • Bring your own model key: Anthropic by default, or any provider (model usage is billed to you by your provider)
Start free

Need more?

Longer runs, more minutes, your company.

  • Longer runs
  • More run-minutes
  • Company accounts
  • Team approvals
  • Your own instance
Talk to us

Questions

What people ask.

Do I need Claude Code?

No. Any script works, including one built on the Claude Agent SDK. A Claude Code project is one way to build a workflow, not the only one.

Which models can it use?

Any model. Anthropic is the default, through your own API key, and you can use any other provider or model with its own key. Model usage is billed to you by your provider.

Where does it run?

In Google Cloud, in the London region (europe-west2). Each run gets its own container.

Can I take it elsewhere?

Yes. It is your repository and a plain YAML file, with nothing proprietary in your code. The runner that runs it is open source (Apache-2.0): foundry-workflow-runner runs the same studio.yaml on your own computer or in a container you build, with no Studio at all. Export to open standards is on the way.

Do you train on my code or data?

No.

What happens if a run fails?

It retries what is safe to retry, then stops cleanly. You get a push, and the run's thread shows exactly what happened.

Your first workflow is one file away.

Free for 3 workflows. Bring your own model key.

Start free