Home › Demo
See It Work

Three runs, start to finish.

Rather than a scripted video, here is exactly what three real AGNT runs do — the trigger, each tool called, and the receipt left behind. Every one of them is buildable in the free tier in about five minutes.

Download and try it Install a ready-made one
Run 01

Something a cloud tool cannot do

This is the clearest illustration of local-first: every step touches a resource that lives on your machine and is invisible to a hosted automation service.

Local repository · runs on your hardware

Review every open branch and write a status brief

Triggered on a schedule each morning. Reads a private repo on local disk, summarizes what changed, and posts to Slack — without the repository ever leaving the machine.

1
Timer trigger fires at 08:00
A scheduled workflow trigger, no external webhook required.
trigger-timer
2
Read local git branches
Shell access to a repository on disk — a private repo that was never pushed anywhere a SaaS runner could reach.
file_system_operation
3
Agent summarizes each branch diff
An agent with repo context in memory reads the diffs and writes a plain-language summary of what each branch is actually doing.
agent · memory: context
4
Post the brief to Slack
OAuth credential pulled from the local vault at call time, never written to a log.
slack
5
Write the receipt
Every step, input, output, and duration persisted to local SQLite.
local receipt
Why it matters: steps 2 and 3 are impossible on a cloud automation platform — the code is not there to read. Platform cost of running this every day for a year: zero.
Run 02

A goal, not a procedure

Here nobody specified the steps. The outcome and the success criteria were given, and the system worked out the rest — then graded its own result and tried again.

Autonomous goal · plan, evaluate, replan

"Find why checkout conversion dropped last week"

One sentence in, an evidence-backed answer out. The plan was generated, not written.

1
Goal decomposed into tasks
Pull conversion metrics, segment by date, isolate the change, correlate with deploys and errors.
goal · plan
2
Query the analytics database
Parameterized SQL against a database bound to localhost.
database_operation
3
Cross-reference deploys and errors
Pulls the release timeline and error spikes to test the correlation, rather than asserting one.
github · sentry
4
Evaluate against success criteria
First pass found the drop but no cause — graded as failing the criteria, not shipped as an answer.
goal · evaluate
5
Replan and run again
A narrower hypothesis, tested against a specific deploy window. Second pass passes.
goal · replan
6
Report with evidence attached
The finding, plus the queries and data that support it, so the conclusion can be checked.
content output
Why it matters: step 4 is the part most systems skip. A goal that cannot meet its criteria says so — it does not narrate a success it did not achieve.
Run 03

Deterministic, with judgment in one place

The hybrid pattern most real automation needs: a fixed path for everything that must not vary, and an agent only at the step that genuinely requires a decision.

Workflow + agent · human approval gate

Triage inbound support mail and draft replies

Runs continuously. Classification and routing are fixed; the reply is written by an agent; sending requires a human.

1
New message arrives
Email trigger on an inbox connected through the vault.
gmail
2
Classify and branch — deterministically
Fixed routing rules. Billing, bug, and feature requests go down different paths every time, with no model variance.
branch
3
Look up the customer record
Reads account and payment history so the reply is grounded in facts rather than guesses.
postgresql · stripe
4
Agent drafts a reply
The one step where judgment helps. The agent's memory carries your tone preferences and past corrections.
agent · memory: preference
5
Human approval before sending
Nothing reaches a customer without a person clicking approve. The gate is part of the workflow, not a policy document.
approval gate
Why it matters: the model never decides routing and never sends. It writes one draft, inside a structure you control, with a person at the end.
Try It

Five minutes to your first run

01 / INSTALL

Download

Windows, macOS, or Linux. No account required, no credit card, no trial clock. Community Core is free permanently.

02 / CONNECT

Add a model

Paste an API key from any of 20+ providers — or point it at Ollama and run entirely offline with no key at all.

03 / RUN

Build or install

Start from a marketplace template and adapt it, or wire your own graph from the node library.

What you will not find here

No sales call, no demo request form, and no gated video. The full product is the free download — installing it is faster than filling in a lead form, and you will learn more.

Download Community Core — free Read the quickstart