Agent Serve · Developers
Write the screens your team needs. Agent Serve gives them sign-in, data, AI agents, security and audit, so you ship domain code only.
Agent Serve, the AI workspace where teams build, deploy, and manage AI agents, lets engineers build small web apps that run inside a workspace. Developers use the agentserve CLI and the app SDK to read tables, query the semantic model, run workflows, hand tasks to Relays and open Chat. Apps run in a sandboxed frame as the person using them, narrowed to their manifest, and deploy from the CLI or from GitHub and GitLab.
How it works
Agent Serve apps follow one path: write the code, upload a draft, and an admin installs it in a workspace.
agentserve app init copies a starter app, its manifest and the app SDK into a new folder.
agentserve app dev serves the app in the same sandboxed frame people see, with your own access.
agentserve app deploy builds and uploads a draft version, or a push to a connected repository builds one.
A workspace admin reviews the permissions the version asks for, makes it installable and installs it.
In depth
Agent Serve gives engineers a CLI, a typed SDK, code tools and Git deploys, so an app is ordinary front-end code.
The agentserve CLI creates, runs, deploys and manages Agent Serve apps, from your machine or from CI.
The Agent Serve app SDK reads and adds table rows, queries the semantic model, runs workflows, assigns Relay tasks and opens Chat.
Ship code tools with an Agent Serve app: TypeScript, JavaScript or Python functions that run on the server, in a sandbox.
Connect a GitHub or GitLab repository, and each push to a chosen branch builds a draft Agent Serve app version.
Security model
An Agent Serve app runs in a sandboxed frame with an opaque origin. Its only way out is the SDK.
Every SDK call runs as the person using the app, authorized again at the time of the call, then narrowed to the manifest.
Agent Serve serves app files under a signed frame token that expires in an hour, and checks the install budget before paid work.
What apps reach
An Agent Serve app lists tables, queries rows and adds rows, with the person’s row security applied.
useTableRows("deals")
An Agent Serve app reads the semantic model and queries measures through it, so numbers match dashboards.
Revenue by region
An Agent Serve app runs deployed workflows, which use the models and integrations already set up in the workspace.
Run "Price check"
An Agent Serve app hands tasks to Relays, the AI coworkers that ask before they act.
Assign to Deal desk Relay
An Agent Serve app opens Chat with a question. The person decides whether to send it.
Ask about this deal
An Agent Serve app can also render dashboard panels and Chat cards, in the same sandboxed frame.
Deal summary card
By the numbers
6
Starter templates in agentserve app init
25 MB
Largest app bundle
10 min
Longest Git build
1 hour
Life of a frame token
Figures are read from the product code.
Before and after
| Work | Before | With Agent Serve |
|---|---|---|
| Sign-in and roles | BeforeBuild a login, sync users and map roles for every internal tool. | With Agent ServePeople open the app already signed in, with their workspace role. |
| Data access | BeforeProvision a database and rewrite access rules by hand. | With Agent ServeThe app reads tables and the semantic model as the person, with row security. |
| AI agents | BeforeWire model keys and prompts into each app. | With Agent ServeRun workflows, assign Relay tasks and open Chat from the SDK. |
| Release | BeforeHosting, secrets and an approval process to set up first. | With Agent ServeDeploy from the CLI or a push. An admin installs the version. |
Under the hood
| Manifest | agentserve.app.yaml: pages, permissions, budget and what to create on install. |
|---|---|
| Bundle | Static files only: HTML, JavaScript, CSS, images, fonts and media. |
| Frame | sandbox="allow-scripts", connect-src none, no inline scripts. |
| Permissions | Tables, workflows, Relays, files, the semantic model and Chat, named per resource. |
| Code tools | TypeScript, JavaScript or Python, run on the server in a sandbox. |
| Git | GitHub App or GitLab token; Bun, npm, pnpm or Yarn lockfiles. |
| Roles | Write uploads drafts. Admin makes a version installable and installs it. |
| Budget | Monthly per install; paid calls stop when it is spent. |
Works with
Apps read and add rows, and installs create the tables an app needs.
Apps query governed measures, so numbers match dashboards.
Apps run deployed workflows and can call code tools from them.
Apps hand tasks to Relays, and Relays can call an app’s code tools.
Permission groups, audit and budgets apply to every app.
Questions
An Agent Serve app is a small web front end that runs inside a workspace. It uses the workspace’s people, tables, semantic model, workflows, Relays and Chat through the app SDK.
No. Agent Serve serves the app’s static files and provides sign-in and data. Logic that must run on the server ships as code tools, which Agent Serve runs in a sandbox.
No. Every call runs as the person using the app and is narrowed to what the manifest names. A manifest never grants access the person does not already have.
Run agentserve app deploy from your machine or CI, or connect a GitHub or GitLab repository so each push builds a draft version. A workspace admin then installs it.
The Agent Serve documentation covers the quickstart, the manifest, the app SDK, the security model, the CLI, code tools and Git deploys, under Build on Agent Serve.
Last updated
Enterprise access
Agent Serve is available to teams by invitation. Tell us what you want to build and we will set up your workspace with you.