Pulse
Surface relevant briefings through the CLI, console, and configured channels. Stay informed about the work your agents are doing.
Read moreOpen source · Active pre-releaseYour intent.
Your intent.
In motion.
An open-source operating system for agents with memory, tools, and authority. From your first intent to a fleet of ongoing work, you stay in control.
SELF-HOSTED · MODEL-AGNOSTIC · MIT LICENSED
+GOVERNED AGENT CORESYS.Z
Autonomous, under your authority.SCROLL TO EXPLORE
At a glance
One lean Go core
Daemon, CLI, embedded console.
Your choice of intelligence
Cloud models or a local brain.
Your rules. Every action.
Governed, journaled, explainable.
01 / INSIDE THE OPERATING SYSTEM
The work is live.
The view is yours.
A console built around the operator's job: talk, observe, automate, govern, and connect. Explore a few of the ways AGEZT makes its work visible.
agezt / consoleINTERACTIVE PRODUCT TOUR
OBSERVE / MONITOR
Mission control
Active runs03Across the sample fleet
Awaiting approval01Your decision comes first
Evidence attached05Inspectable sample events
Run activity
Workspace reviewagent / reviewer
WorkingResearch digestagent / researcher
VerifyingScheduled briefagent / daily-brief
WaitingEvent stream
Intent received00:00
Model route resolved00:01
Capability allowed00:02
Workspace inspected00:03
Result recorded00:04
Illustrative interface study · sample data · no live daemon connection36 VIEWS / 8 OPERATOR SECTIONS
01 / 04
Watch the work, not just the answer.
Follow running tasks, delegated agents, tool decisions, and their shared event trail. The live console lets you pause or steer a run while it is in flight.
Explore the console02 / A SYSTEM THAT ACTS
Beyond the
chat window.
Give your agents a place to work, a memory to build on, and boundaries to respect. AGEZT connects the entire loop.
THINKER → WORKER → VERIFIERF.01
A fleet. Not a conversation.
Durable agents with their own identity, memory, and lifecycle. Delegate work, coordinate through a shared mailbox, and keep the whole fleet in view.
agezt / core
Researcher Gathering context
Builder Executing plan
Verifier Checking evidence
YOU SET THE LIMITSF.02
Autonomy with a boundary.
Edict checks each tool call. Allow it, require your approval, or deny it. Pause everything with a single command.
Workspace readAllow
External actionAsk you
Outside scopeDeny
$ agt halt■
KNOWLEDGE, CONNECTEDF.03
Context that stays.
Persistent memory, a world graph, and reusable skills. Each run starts with more than a blank prompt.
PROOF OVER PROMISESF.04
The receipts are built in.
Every step joins a hash-chained journal. Follow a decision back to its intent, replay a run, and see why it happened.
$ agt why <event_id>intenta7f2
policyc09e
actionb4d1
proofe83a
BLAKE3 / hash-chained
Durable work
Typed workboard tasks, acceptance criteria, and OKR roll-ups connect agent activity to work you can actually account for.
Read moreGoverned extension
Connect MCP servers, install capability packs, or promote a tested tool through Forge with operator approval.
Read more03 / PUT INTENT TO WORK
One system.
Many ways to work.
A workspace review. A deeper investigation. A recurring brief. The same governed runtime supports very different kinds of work.
FOR DEVELOPERS
Give complex work a team.
Use the Conductor to split a difficult task across Thinker, Worker, and Verifier roles. Keep the plan, execution, and checks in one governed runtime.
- 01Understand the workspace and define the task
- 02Coordinate specialist work under scoped capabilities
- 03Check the result and retain an evidence trail
EXAMPLE OUTCOME
A review you can follow
Workspace findings
Proposed improvement
Verification evidence
conductorfiledelegateFROM THE CLI
agt conductor "Review this project and propose a focused improvement"Coding-agent integration is opt-in and returns a diff; it does not merge changes.
04 / FROM INTENT TO EVIDENCE
See the loop.
Feel the control.
A small walkthrough of a big idea: autonomous work should always leave a trail you can follow.
YOUR INTENT
“Review my recent runs and prepare a morning brief.”
Illustrative simulation · no models or tools are called
Intent01
Governor02
Edict03
Tool04
Journal05
EXECUTION JOURNAL—
›Waiting for the first step…
Ready for your intent.Step 0 of 5
05 / THE OPERATING SYSTEM UNDERNEATH
Nothing moves
without a trail.
Many entry points. One governed loop. Follow an intent through routing, authority, action, and an inspectable record.
MANY ENTRY POINTS / ONE RUN
agt run "Explain this project"An operator request, a scheduled wake, an inbound message, or an API call starts the run. A correlation ID connects the work from this point onward.
Every stage joins the journal.BLAKE3 / CORRELATION ID / REPLAY
THE EDICT POLICY ENGINEChoose how far
Choose how far
autonomy goes.
Grant capabilities deliberately. Keep an approval boundary for actions that need you. Make denials as visible as successes.
Understand the authority modelILLUSTRATIVE REQUEST
Read a file inside the workspace
Authority check
file.readThis example is within the agent's granted capability and workspace scope. The action can proceed, and the decision joins its audit trail.
policy.decision → allowed → tool.result06 / CONNECTED, BY DESIGN
Change the brain.
Keep the system.
Pick a provider. Connect your channels. Extend with MCP and plugins. Your operating system stays yours.
MODELS & PROVIDERS14 / brands
Anthropic
OpenAI
Google
Ollama
Mistral
DeepSeek
Z.ai / GLM
MiniMax
Kimi
Qwen
AWS Bedrock
Azure OpenAI
Cohere
Vertex AI
AGEZT
WHERE YOUR WORK HAPPENSExplore integrations
Provider and channel availability depends on your configuration.
Logos identify supported integrations and do not imply endorsement. Matrix logo is a trademark of The Matrix.org Foundation.
Fits into your stack.
OpenAI-compatible endpoints, a native REST API, and official Python, TypeScript, and Rust clients.
BUILT TO BE BUILT ONYour application.
Your application.
The same authority.
Submit work from a client, an IDE, or your own application. The entry point can change while the governed execution path stays connected.
- Streaming runs and inspectable outcomes
- Bearer authentication and tenant-aware clients
- A shared mailbox for agent coordination
request.sh
curl http://127.0.0.1:8800/api/v1/runs \
-H "Authorization: Bearer $TOKEN" \
-H "Content-Type: application/json" \
-d '{"intent":"Explain this project"}'SSE-READY / CORRELATION IDs
Submit an intent to the governed runtime. Add stream: true for SSE, then retrieve the journaled run using its correlation ID.
Enable the corresponding API listener and use the token issued by your daemon. These examples are documentation; this site does not execute them.
07 / EXPLORE THE OPEN SYSTEM
Go deeper.
Build something yours.
The source, the guides, and the interfaces are open. Find a practical next step, whether you are operating AGEZT or building on it.
Open source is part of the architecture.
Read the sourceMIT licensed. Self-hostable. A readable Go core with an explicit dependency policy. Examine the implementation and its current limits in the repository.
08 / MAKE IT YOURS
Your machine.
Your authority.
Start from the source. Choose your intelligence. Give your first intent a place to become action.
Explore the sourceActively developed · see repository for release status
BUILD & QUICKSTART
git clone https://github.com/agezt/agezt.git
cd agezt
npm --prefix frontend ci
npm --prefix frontend run build
go build -o agezt ./cmd/agezt
go build -o agt ./cmd/agt
./agt quickstart
01
Build your runtime
Compile the Go daemon and operator CLI, with the embedded console built from the frontend.
agezt + agt02
Choose your intelligence
Sync the catalog, configure a provider, and pick its model. Quickstart prints the start command for your setup.
agt quickstart03
Run. Inspect. Refine.
Check the daemon with doctor, submit an intent, and use the journal to understand the work it performed.
agt doctor → agt run → agt whyAn open-source agentic operating system: a Go daemon, an operator CLI called agt, and an embedded React console. It coordinates agents, models, tools, memory, and automation through one governed runtime.
Yes. AGEZT supports nine adapter families, including OpenAI, Anthropic, Google, Ollama, and AWS Bedrock. The model catalog drives provider configuration. There is no default provider or model: you choose one, or configure per-task routing.
Every tool call is checked by Edict policy and every step is journaled. You decide which capabilities are allowed, which need approval, and which are denied. The agt halt command freezes execution immediately.
Yes. AGEZT is MIT licensed and runs on your own machine. Provider API costs and third-party service requirements depend on the integrations you choose. See the repository for the current implementation status and known limits.
An agent is a named identity with its own lifecycle, memory, skills, mailbox, workspace, and authority. It can sleep and wake again instead of being reduced to a single prompt or a temporary chat session.
Registered tools are gated by capabilities and Edict policy. The policy can allow the request, require operator approval, or deny it. Decisions and results join the run's journal so you can inspect what happened.
Ollama is a supported adapter family. Discover the local models, choose the provider and model explicitly, and keep the runtime on your own machine. Local model features depend on the model and your hardware.
Tools declare a rollback class. Recovery depends on the specific tool and the action it performed; an audit trail does not make an irreversible external effect reversible. Check the operator documentation for supported recovery behavior.
A schedule is typed infrastructure that can wake an agent, run a workflow, execute a system task, or invoke a governed tool. Every firing is recorded, so autonomous work can be traced back to its trigger.
AGEZT is actively developed and its README currently identifies it as a pre-release. Core capabilities are implemented and being hardened. Review the current release notes, operator guides, and known limits before choosing a deployment.
Intent is just the beginning.
Build what happens next