Build agents you trust
Connect your tools. Set the rules. Build and run AI agent teams with built-in guardrails, human approvals and complete visibility. Free, open source and self-hosted.
npx agentbuildkitApache 2.0. macOS and Linux, Node.js 18 or newer. The demo and dashboard need no API key.
Manager
Account
Service
Reviewer
Local dashboard with demo data. Account A-100 is a fixture.
Connected apps through Composio. Google Sheets today, read-only. Each connected tool has to be explicitly approved, is pinned to a version and only reads, so more apps can be added one at a time. More are on the roadmap.
One command. Running in minutes.
It checks your system, installs AgentBuildKit and both templates in their own environment, checks they work, then offers the dashboard and the offline demo.
macOS or Linux and Node.js 18 or newer. Windows hasn’t been verified for the dashboard.
The agentbuildkit 0.2.0 harness, the agentbuildkit-templates 0.2.0 package and the agentkit command, using uv. Python 3.12 is set up for you.
An isolated uv tool environment. It never changes your system Python. If uv is missing, it asks first.
Not needed for the demo or the dashboard. For live runs, set OPENAI_API_KEY or ANTHROPIC_API_KEY.
agentkit dashboard runs on 127.0.0.1 for one local operator, with no login. Opening it never calls a model.
uv tool uninstall agentbuildkit
Every run, every approval, in view
The local dashboard shows what your team did, what it’s waiting for, and what needs a check. Approve or reject changes from the browser. It reads what was recorded and never invents a step.
Sent at most once. Never replayed.
Nothing is sent.
Approve, with the facts in front of you
Demo dataThe action, the customer and what happens either way. Nothing is sent until you decide.
See the route each request took
Demo dataSteps the run recorded get a blue ring. Agents it didn’t use stay faint.
An audit trail you can read
Demo dataEvery agent run, tool call, pause and approval, with the provider and model. Hashes, never argument values.
Screens use the shipped demo data. One local operator, bound to 127.0.0.1. Opening the dashboard doesn’t call a model.
Start from a working team
Two templates ship with agentbuildkit-templates. Create your own instance, edit its draft, evaluate it offline, then activate it.
Customer Service
Approvals onAlex classifies a request. Python routes it to Sam or Taylor. Jordan reviews the reply. A service change pauses for your approval.
Operations Reporting
Read-onlyAn analyst reads the jobs and writes a summary. A reviewer checks it names only those jobs. No approval actions, nothing written back.
- 1Create teamAn inactive instance of a template.
- 2Edit the draftVersioned. Nothing goes live.
- 3ValidateChecked against the template.
- 4Evaluate offlineScripted suite, no API key.
- 5ActivateExplicit. Rollback is too.
Draft teams from Claude or Cursor
A local MCP server can create an inactive team, edit its draft, validate it and run the scripted evaluation. It can’t activate, run, approve or connect anything: that stays with you in the dashboard.
Guardrails are part of the architecture
The production layer around your agents, built on Microsoft Agent Framework. Plain Python, not the model, decides which tools an agent may use and whether a side effect may run.
Agent teams
Typed agents with clear roles. An orchestrator routes work to specialists and a reviewer checks the result.
Your model, per agent
OpenAI and Anthropic built in. One workflow can mix them without branching on the provider.
Deterministic workflows
Plain Python decides the route, not the model. Unknown routes fail closed.
Tool allowlists
Each agent gets only the tools you grant. Anything else is rejected and logged.
Human approval
Reads run. Side effects wait for a person, and an approved action runs at most once in the local store.
Durable state and audit
Runs pause and resume from SQLite. Every decision is recorded with its provider and model. OpenTelemetry export included.
Offline evaluation
Scripted suites check the route, tools and approvals against expectations, with no API key.
Tools and connectors
GitHub, Cursor, MCP servers and your HTTP APIs, each behind a permission. Optional read-only Google Sheets via Composio.
A CLI that tells you the truth
agentkit doctor checks Python, SQLite, templates and which providers are configured.
agentkit demo --approve
Your keys, state and audit records stay in your environment. Only the prompts you send to your chosen model leave it.
Where AgentBuildKit is going
The framework stays free. These are planned, not shipped, and none has a date or a price.
Persistent agent memory
Context that carries across runs, without stuffing everything into the prompt.
Skills
Reusable capabilities an agent can be granted, under the same allowlists and approvals.
Ready-to-deploy teams
Complete agent teams for real business outcomes, shared through a template store. None are on sale yet.
More connected apps
Read-only connections to more apps through Composio, added and tested one at a time, starting with the ones teams use most.
Hosted, multi-user deployment
Today it’s one local operator. A hosted option with several users is on the roadmap.
Questions
Is it free?
Yes. AgentBuildKit is open source under the Apache 2.0 licence. You pay only your model provider, and only for live runs.
Do I need an API key?
Not for the offline demo, the dashboard or the evaluation suites. Live runs need an OpenAI or Anthropic key for the agents that use them.
Which models does it support?
OpenAI and Anthropic are built in, and one team can mix them. Microsoft Foundry is an optional extra.
Where does it run?
On your machine. The dashboard binds to 127.0.0.1 for one local operator. It isn’t a hosted or multi-user service.
Does it work on Windows?
The bootstrapper supports macOS and Linux. Windows hasn’t been verified for the dashboard or the MCP server.
Can I build my own team?
Today you create instances of the two shipped templates, edit their drafts, evaluate them offline and activate them. You can also build agents directly in Python with the harness.
Does approval guarantee the action happens once?
Approval runs an action at most once in the local store. An interrupted action stays unresolved and is never replayed. It isn’t exactly-once delivery to an outside system.
What is it built on?
Microsoft Agent Framework. AgentBuildKit adds the guardrails, approvals, state, audit and evaluation around it.
Install it now
Run the demo and open the dashboard in a few minutes. No account, no API key, nothing to pay.
npx agentbuildkit