One command sets it all up, then offers to start it for you.

It asks you a few questions, then builds the project and sets up a password for your dashboard. When it finishes it prints the last three commands to run. Turn everything off and back on whenever you like, because your agents carry on from where they stopped.

People run things like: draft the release notes, send the deploy summary, summarize yesterday's error logs.

agent: :2000
dashboard: :4000
postgres: :5433

evestack

npx evestack create

… asks a few questions, saves your settings, installs everything

docker compose up -d postgres

✓ evestack-postgres-1 | database system is ready to accept connections

npm run db:bootstrap

✓ Schema created.

npm run dev

✓ eve dev ready on http://localhost:2000

✓ [world-postgres] Re-enqueued 2 active run(s) on startup

evestack

self-hosted

Sessions

33

Turns

119

Tokens in/out

182K/53K

Model spend

$0.06

Infrastructure

$0.00

SessionStatusTurnsModelInOutCostStarted

Which sessions failed in the last hour?

None. Every session completed cleanly. The longest ran 41.0s with 13 tool calls.

What did the deploy email session cost?

Deploy summary email ran 3 turns with 5 tool calls in 18.2s, costing $0.0034 in model spend. Infrastructure: $0.00.

Everything you need, already wired together

Agents that survive a restart, run code safely, work to a schedule, and ask before they do anything you would not want.

Nothing gets lost
Every step is saved to your database the moment it happens. Shut it down mid conversation, or reboot the machine with agents halfway through a job, and they pick themselves back up where they stopped.
35 run_created
36 step_started
37 ai.streamText
38 step_completed
No hosting bill
The database, the sandbox and the dashboard all run on your machine, so there is nothing to bill. The only thing you pay for is the AI model, and you can run that locally too.
$0.00infrastructure / month
Your prompts stay private
Flip one setting and the words in your conversations never leave the agent. You still get timings and token counts, just not the content.
EVESTACK_TRACE_CONTENT=off
prompt: ••••••••••••
result: ••••••••
See every step it took
Open any conversation and follow exactly what happened: each turn, every tool the agent reached for, and every call it made to the model, with what went in and what came back.
agent.session
└ agent.turn
└ agent.step
└ ai.streamText
└ ai.streamText.doStream
It asks before it acts
Mark any tool as needing permission. The agent stops and waits while you approve or deny it in the dashboard, and every decision is recorded.
send_email wants approval
approvedeny
Runs while you sleep
Give an agent a schedule and it works on its own. Every run is kept, so you can see what happened at 3am, and you can pause one without taking anything else down.
heartbeat0 * * * *
last run 41m ago ok
next in 19m

Watch your agents. Step in when it matters.

Every conversation your agents have, read straight out of your own database. What they did, what it cost, how long it took, and when a run goes wrong you can turn that exact run into a test.

evestack

Overview

Sessions

Chat

Integrations

Observability

Monitors

Traces

Observability/Monitors

ProductionLast 12 hoursself-hosted

Runs

Error 0%

Session duration

p5015.8sp7520.8sp9536.6sp9940.1s

41.0speak

Session

Tokens

Duration

Cost

Write a detailed 1500-word essay about the history of database indexing.

8,827

41.0s

$0.0077

Use your agent tool to delegate this subtask to a subagent…

15,926

28.4s

$0.0021

Deploy summary email to the team

15,011

18.2s

$0.0034

Draft release notes for v0.4.0

13,531

16.9s

$0.0031

Show 101 of 1

read from your own database, 100% yours

built byapp/sessions/page.tsx app/sessions/[id]/page.tsx lib/pricing.ts app/approvals/page.tsx

It waits for you

Mark a tool as risky and the agent will not use it until you say so. Approve or deny it in the browser, and the decision is written down.

send_email

{ "to": "team@…", "subject": "Deploy done" }

requested

send_email

{ "to": "team@…", "subject": "Deploy done" }

approved

send_email

{ "to": "team@…", "subject": "Deploy done" }

executed

Connect the tools you already use

Sign in once from the dashboard and your agent can use Gmail, GitHub, Slack, Notion, Linear and about a thousand more.

agent:2000

Four pieces, all on your machine

The agent does the work. Postgres remembers every conversation. Docker keeps any code the agent runs inside a box where it cannot touch the rest of your computer. The dashboard shows you all of it, and none of it talks to the outside world.

Here are the actual files that do it, read straight from the repository.

docker-compose.ymlThe whole setup is one Docker file

name: evestack

services:
  postgres:
    image: pgvector/pgvector:pg17
    environment:
      POSTGRES_DB: ${POSTGRES_DB:-evestack}
    ports:

  dashboard:
    build:
      context: .
    depends_on:
      postgres:
        condition: service_healthy

agent/agent.tsSessions saved to Postgres, talking to the model directly

import { openai } from "@ai-sdk/openai";
import { defineAgent } from "eve";

const workflow =
  process.env.WORKFLOW_POSTGRES_URL
    ? { world: "@workflow/world-postgres" }
    : undefined;

const model = openai(
  modelId,
);

export default defineAgent({
  model,
  experimental: { workflow },
});

agent/instrumentation.tsTen lines, and the dashboard can see everything

import { registerOTel } from "@vercel/otel";
import {
  defineInstrumentation,
} from "eve/instrumentation";

export default defineInstrumentation({
  setup: ({ agentName }) => {
    registerOTel({
      serviceName: agentName,
      traceExporter:
        new OTLPHttpJsonTraceExporter({
          url: endpoint,
        }),
    });
});

docker-compose.yml

name: evestack

services:
  postgres:
    image: pgvector/pgvector:pg17
    environment:
      POSTGRES_DB: ${POSTGRES_DB:-evestack}
    ports:

  dashboard:
    build:
      context: .
    depends_on:
      postgres:
        condition: service_healthy

The whole setup is one Docker file

agent/agent.ts

import { openai } from "@ai-sdk/openai";
import { defineAgent } from "eve";

const workflow =
  process.env.WORKFLOW_POSTGRES_URL
    ? { world: "@workflow/world-postgres" }
    : undefined;

const model = openai(
  modelId,
);

export default defineAgent({
  model,
  experimental: { workflow },
});

Sessions saved to Postgres, talking to the model directly

agent/instrumentation.ts

import { registerOTel } from "@vercel/otel";
import {
  defineInstrumentation,
} from "eve/instrumentation";

export default defineInstrumentation({
  setup: ({ agentName }) => {
    registerOTel({
      serviceName: agentName,
      traceExporter:
        new OTLPHttpJsonTraceExporter({
          url: endpoint,
        }),
    });
});

Ten lines, and the dashboard can see everything

Run it yourself, or pay someone to run it

Both options use the same open source framework underneath. It is called eve, and Vercel builds it. The difference is whose computer your agents run on, and who can see what they are doing.

A hosted service

evestack on your hardware

Runs on
A hosted service
Someone else's servers
evestack on your hardware
Your machine, server, or cluster
Where conversations are stored
A hosted service
Their platform
evestack on your hardware
A Postgres database you own
How long history is kept
A hosted service
Set by the provider
evestack on your hardware
As long as you keep the rows
Dashboard
A hosted service
Included, and it watches
evestack on your hardware
Included, and it can also act
Who can reach your data
A hosted service
You and the provider
evestack on your hardware
You
Setup
A hosted service
Deploy to their platform
evestack on your hardware
One command, then four more

	A hosted service	evestack on your hardware
Runs on	Someone else's servers	Your machine, server, or cluster
Where conversations are stored	Their platform	A Postgres database you own
How long history is kept	Set by the provider	As long as you keep the rows
Dashboard	Included, and it watches	Included, and it can also act
Who can reach your data	You and the provider	You
Setup	Deploy to their platform	One command, then four more

Your agents, on your hardware.

Five commands and it is running on your machine.

npx evestack create

Star on GitHub

Or read the documentation first. It covers how the pieces fit together, how to run this on a server rather than a laptop, and how to fix it when something breaks.

Run AI agents on your own machine.

One command sets it all up, then offers to start it for you.

Everything you need, already wired together

Nothing gets lost

No hosting bill

Your prompts stay private

See every step it took

It asks before it acts

Runs while you sleep

Watch your agents. Step in when it matters.

It waits for you

Connect the tools you already use

Four pieces, all on your machine

Run it yourself, or pay someone to run it

Runs on

Where conversations are stored

How long history is kept

Dashboard

Who can reach your data

Setup

Your agents, on your hardware.