Quick Start

Ten minutes gets you from a blank chat to a downloaded file that AI Partner researched and built on its own. Then another ten tells it who you are, which is what turns it from a tool into a partner.

If you're reading this on the hosted site, you already have everything you need — just sign in. Self-hosting the free V1 edition instead? Install it from the GitHub repository, then rejoin this guide at Step 2 — the app itself is the same from there on. See the editions.

Step 1: What you need

One model key

OpenRouter is the quickest — free tier, one key, 100+ models. Sign up at openrouter.ai.

Nothing else

No GPU. No local model. Everything runs against cloud models, and you can add or change providers whenever you like.

Step 2: Let the setup wizard run

The first time you open AI Partner, a short Setup Wizard appears and asks for three things:

1
A model provider
Pick where your models come from. You can add more later, and route different kinds of work to different models — see Model Routing.
2
Your first key
Paste it in. It's stored encrypted and is never included in prompts sent to model providers.
3
Your name and role
Used straight away, in every conversation. You'll expand on this in Step 5.

You can skip the wizard and do all of it later from Settings → Models. You can also re-run it any time from System → Re-run Setup.

Provider	Where to get a key	Worth knowing
OpenRouter	openrouter.ai	Free tier, 100+ models behind one key — easiest start
Anthropic	console.anthropic.com	Claude models
OpenAI	platform.openai.com	GPT models
Groq	console.groq.com	Free tier, very fast

Some features need a model that can read an image — screen control and vision-based web navigation among them. If you only add a text-only model, those features say so plainly rather than failing oddly.

Step 3: Run your first goal

Type this into the chat box, exactly as it is:

Research the 5 most-starred GitHub repositories in the last 7 days.
Write a summary with each repo's name, stars, language, and a one-line
description. Then generate an Excel file with the results.

Click Send as Goal — the rocket icon — rather than plain send.

Then watch the Goal Progress panel on the right:

1
It plans
The goal is broken into steps: research, extract, build the file.
2
It works
Each action appears as it happens — searching, fetching pages, processing the data, building the workbook.
3
It checks itself
Before calling it done, it confirms the file exists and actually has data in it.
4
You get the file
A download appears in Work → Generated Files.

First run takes 30–90 seconds.

No goal panel? You probably used plain send. Chat is for quick questions and answers; Send as Goal runs the full autonomous loop. The difference is explained in Goal Execution.

Step 4: Ask it what else it can do

Type "what can you do?" It answers from its own current configuration — which models are connected, which services are linked, whether a computer is connected — not from a generic feature list. It's the fastest way to find out what's live on your instance.

Step 5: Tell it who you are

This is the step people skip and then wonder why the agent feels generic. Everything proactive — briefings, meeting attendance, judgement about what matters — depends on it.

Open the Workspace browser in the left sidebar and edit these four files. The editor always points at your own workspace, so this is safe on a shared instance.

USER.md — who you are

Included in every conversation. Keep it short and factual.

# User Profile
Name: Alex Chen
Role: Founder & CEO, Acme Inc.
Current focus: Closing Series A, Q2 2026
Key relationships: Sarah (CFO), Mike (CTO), Priya at Sequoia
Communication style: Direct. No filler. Lead with data.

GOALS.md — what you're working toward

Meeting attendance, proactive work and weekly briefings all align to these automatically.

## Active Goals

### Close Series A
- Success: Term sheet signed by June 30, 2026
- Context: In conversations with Sequoia, Accel, and Tiger Global

### Launch v2.0
- Success: 1,000 active users by July 15, 2026
- Context: Feature freeze April 30; beta starts May 15

HEARTBEAT.md — what to watch without being asked

Standing work that runs on your schedule, with no prompt from you.

## Standing Tasks

- Every morning (7 AM): Summarize top 5 tech and startup news stories
- Every weekday (9:15 AM): Check NIFTY 50 and my portfolio stocks
- Every Friday (5 PM): Summarize this week's meetings and key decisions

AUTHORITY.md — what it may do without asking

The most important file here. It decides what happens automatically, what comes to you for a one-tap approval, and what is never allowed at all.

## Auto (do it, and log it)
- action: read_.*
- action: send_message + counterparty: family

## Draft + ask (send me an approval request)
- action: reply_email + counterparty: colleague
- action: reply_email + counterparty: client
- action: send_slack_dm + counterparty: colleague

## Block (never, no exceptions)
- action: payments.*
- action: legal_commitment
- action: place_phone_call

Full authority policy guide →

Step 6: Connect a channel so it can reach you

Approvals, briefings and finished files are far more useful on your phone than in a browser tab you've closed. System → Messaging connects Telegram, WhatsApp, Slack, Discord or Signal — and it's the same partner with the same memory, wherever you talk to it.

Messaging channels →