Overview

AI agents

A service on your project canvas that answers with your instructions, with a team chat, an HTTP API and triggers that start chats on their own.

Coming soon
AI agents are not available on StackBlaze yet. This page describes how they work so you can plan for them; names, limits and details can change before they are released.

An AI agent is a service on your project canvas that answers with your instructions. It has a team chat, an HTTP API, and triggers that start chats on their own: a schedule, a webhook, or a chat bubble on your website.

You set it up in the dashboard. There is no code to write and no image to build. The agent runs in its environment like any other service.

These are not the same as the StackBlaze AI Agent, which operates StackBlaze itself for you. An AI agent is something you build and run for your own users and team.

The parts

PartWhat it is
InstructionsWhat the agent does and how it answers. Up to 20,000 characters.
ModelThe model it uses. By default the agent gets its own AI Gateway, which holds your provider key. The agent never holds that key.
ToolsWhat the agent may do while it answers: read and search its files, search the web, keep a task list. Shell and file-editing tools are off unless you turn them on.
KnowledgeFiles the agent reads and searches: help articles, policies, exports.
SkillsStep-by-step notes for one kind of task, read when a request matches.
IntegrationsOutside services the agent uses as tools, each a hosted MCP server.
TriggersTime (schedules), Webhook, API keys and the Website widget.
SessionsEvery chat the agent had, from every channel, with its transcript.
EvalsRepeatable tests that check whether quality holds after a change.
OptimizationsCheaper ways to run the agent, measured against its own recent sessions.

Each part is a section tab on the agent's Agent tab: General, Knowledge, Skills, Integrations, Triggers, Evals, Optimizations, Sessions. Every section title has an info icon that explains the section.

What it runs on and what it costs

The agent is a service with a 1 GB volume for its files, sessions, evals and optimization runs. Its chat history is kept in PostgreSQL. With its own AI Gateway, the agent and the gateway share one PostgreSQL and show as one app on the canvas. Both run in your environment and use resources like any service; see Pricing.

Model calls go through the gateway to your provider with your key, so your provider bills the tokens. With an AI model of the same environment, the agent uses that model instead.

Every chat uses tokens: team chats, API calls, webhooks, schedules, widget visitors, eval trials and optimization runs. Sessions show the tokens in and out of each chat.

Quick start

  1. 1

    Add the service

    Open a project and add a service on the canvas. Pick AI agent.
  2. 2

    Pick a template

    Under Start from, pick a template (Customer support, FAQ bot, Internal helpdesk, Webhook triage, Daily report) or start blank.
  3. 3

    Choose the model

    Under Model, pick a provider and paste its API key. Deploy also deploys an AI Gateway for this agent, in about two minutes. The key is kept in the gateway, not on the agent.
  4. 4

    Deploy

    Click Deploy.
  5. 5

    Give it knowledge and talk to it

    Open the agent from the canvas. Upload files in Knowledge, then use Open chat on its card to talk to it.
  6. 6

    Call it from your code

    Create a key in Triggers → API. See Triggers.

In this section

PageWhat it covers
Create an agentTemplates, the model and its AI Gateway, instructions, tools and updates.
KnowledgeFiles the agent reads and searches.
SkillsSteps for one kind of task, read when a request matches.
IntegrationsHosted MCP servers as tools.
TriggersSchedules, webhooks, API keys and the website widget.
SessionsEvery chat, by channel, with its transcript.
EvalsTest cases, checks and results on three axes.
OptimizationsCheaper setups tried against recent sessions.
APIStackBlaze routes for agents and the agent’s own chat API.
Security and limitsKeys, the gateway, public agents and limits.

To operate StackBlaze itself from an editor or terminal agent, see the AI Agent and the MCP server.