AI agents & MCP

AI that works inside your rules.

Build agents with a personality, instructions and the tools they may use. Connect any AI client over MCP. Let an agent configure the tenant or build your apps. Every one of them acts under a real identity, with real permissions.

AI agents documentation

01PersonaAva · Sales assistant

02Identityava@acme.com · role Sales

03Toolsrecords · calendar (MCP) · web

04ModelClaude · 30 tool steps

05Instructions“Be brief. Never promise discounts.”

Anatomy of an agent

Five layers. Only one of them is the prompt.

An agent is never defined by its prompt alone. The prompt says what role to play; its user, its kind and its tools decide what it can actually do.

01
Persona
The name and face people meet in chat, and who may see it.
02
Identity
A dedicated user. Its roles are the hard permission boundary.
03
Tools
The tenant's tools it may use, narrowed per agent, plus external MCP servers.
04
Model
Claude, GPT or Gemini, with your key and limits on output and tool steps.
05
Instructions
How it behaves: its tone, its rules, what it must never do.

Assistants and builders

Agents that do the work. Agents that build.

Assistants

Work with records: answer questions, summarize, draft and create follow-ups. The data model is read-only for them, whatever their roles.

Business work inside your apps
A read-only data model
Writes automatic or confirmed

Builders

Change the application itself through the MCP builder: the data model, roles and permissions, forms, automations, views and apps.

Configure the tenant on development
Build apps in the Studio
Every change checked before it applies

Beyond your tenant

Tools from wherever you work.

Agents use the tenant's own tools and any external MCP server you connect: a calendar, a document store, a code host. They can call outside web APIs, read the saved responses and keep notes as they go.

01External MCP serversConnect any server built on the MCP standard and choose which of its tools each agent may use.

02Web requestsCall outside APIs during a conversation and work with the saved responses.

03NotesAn agent keeps notes in the conversation, so long tasks don't lose their thread.

Human in the loop

It asks before it writes.

With confirm-before-apply, every call that isn't a read waits for a person. The card shows exactly what will happen, including a check of any configuration change. Decisions are stored with the conversation.

Ava

Sales assistant

Personality: Warm, brief, never promises discounts.
Model: Claude
Acts as: ava@acme.com · role Sales
Tools: CRM recordsCalendar (MCP)Web requests
Writes: Asks first

run · lead #1042working

MCP, both ways

Your tenant, in every AI client.

Every tenant serves an MCP server for records and an MCP builder for configuration. Claude, ChatGPT, Cursor or Claude Code sign in with OAuth, two-factor included, and work with that person's permissions.

1# Records, with your permissions (sign-in opens in the browser)
2claude mcp add --transport http acme https://acme.jetstack.app/api/mcp
3
4# Configuration: roles, permissions, forms, automations, apps
5claude mcp add --transport http acme-builder https://acme.jetstack.app/api/mcp/builder

MCP builder

Anything you can click, an agent can configure.

The MCP builder exposes the tenant's configuration as documents an assistant reads and changes, always in two steps: a check that shows every effect and diff, then the apply. Every applied change lands in the event log and the change history.

Errors point to the fix: a misspelled name comes back with the closest names that exist.

Roles
Access & field grants
Permission queries
Users
API keys
Secrets
Environment values
Properties
Forms
Record actions
Record tabs
Automations
Schedules
Filter lists
Tags
Queries
Views
Canvases
Modules
Apps

Where agents work

In the back office. In every app.

Back-office chat

Next to every screen, aware of the record or selection on it.

Inside your apps

Embed the chat in any page: one agent, its own conversations, the page's context.

In automations

Prompts and agent runs as steps, with structured JSON answers.

Over REST

Start prompts and agent runs from your own services and follow them live.

Models

Your models. Your keys.

Claude, GPT, Gemini

Choose the model per agent; the catalog includes the Claude 5 family.

Your provider keys

Keys live in the secret store, and your provider bills the usage.

Limits per agent

Output tokens, tool steps and key lifetime, set for each agent.

Prompt caching

Long conversations re-read from the provider's cache at a fraction of the price.

AI FAQ

AI on Jetstack, answered.

Can an agent do something it shouldn't?

An agent can only do what its own user may do. Assistants and builders are separate kinds, writes can wait for a person's confirmation, and every step is recorded in the conversation and the event log.

Which AI clients work with the MCP server?

Any client that supports MCP with OAuth sign-in, such as Claude, ChatGPT, Cursor and Claude Code. Local assistants can also connect through the jetstack CLI.

Can agents use our other systems?

Yes: through connectors, external MCP servers and web requests, each enabled per agent.

Does the MCP builder work on production?

Under the production change policy you choose. A locked production refuses configuration changes from anywhere except a promotion from development.

Put AI to work

Give an agent its first job.

Tell us which job you would hand to an agent first. We'll show you how it would run, and where it would ask before acting.

A first agent on your data
The tools and permissions it needs
Where a person stays in the loop