What Is AI Partner?
Not a chatbot. A partner.
Most AI tools are glorified search bars: you ask, they answer, the conversation ends and nothing persists.
AI Partner is different. Give it a goal — "research our top five competitors and produce a slide deck" — and it decomposes that goal into steps, runs them autonomously (web search, data synthesis, PowerPoint generation), monitors its own progress, corrects its mistakes, and delivers a file. You don't babysit it.
Go further: give it access to your email, Slack, Telegram, and calendar. It reads your incoming messages, drafts replies in your writing style, and sends them only after you approve via a single Telegram button-tap. It joins meetings you can't attend, listens and speaks as you, and hands you a transcript and action items when it's done.
And it doesn't wait to be asked. A background loop runs on your schedule — morning market briefings, competitor news checks, weekly summaries — and pushes results to your preferred channel without any prompt from you.
Clone, add one API key, type a goal. Under 10 minutes.
Six annotated walkthroughs of real tasks from start to finish.
The ReAct loop, authority gating, and memory system explained simply.
Email, Slack, Telegram, phone — acting as you with your approval gate.
The difference at a glance
| Chatbot / LLM wrapper | AI Partner | |
|---|---|---|
| Input | Prompt → response | Goal → multi-step execution plan |
| Loop | Single turn | ReAct loop: reason → act → assess → replan |
| Tools | Optional function calls | 160+ tools across 41 connected services, plus a safe place to run code |
| Memory | Context window only | Your history, your facts, your documents and everyone you deal with |
| Output | Text | Files, reports, sent messages, standing jobs |
| Recovery | None | Corrects itself, notices when it's stuck, asks you when it should |
| Autonomy | Waits for prompts | Acts on your schedule without being asked |
| Human proxy | None | Joins meetings, replies to email/Slack/Telegram in your voice, takes phone calls |
| Your computer | None | Works on your own machine — your files, your apps |
| External agents | None | Agent Mesh — discover, delegate, classify, and audit untrusted internet agents |
The key distinction: every other tool resets when the conversation ends. AI Partner accumulates context across weeks, holds goals across quarters, attends meetings you aren't in, and reports outcomes against criteria you defined — without being asked each time.
Three editions: V1, V2 and V3
Each edition is a superset of the one before it. Pages documenting an edition-specific feature carry a badge at the top.
| V1 — works for you | V2 — works as you | V3 — works with you | |
|---|---|---|---|
| What it is | The free, self-hosted agent, for one person | The hosted, invite-only edition — acts on your behalf, multi-user and governed | The partner edition — reaches your own machine, keeps standing work alive, and holds a relationship with you |
| Access | Free — self-hosted | Invite-only, available now | Invite-only, available now (this hosted site is V3) |
| Core (in all three) | Goal execution with checked outcomes, memory and your own documents, file generation, messaging channels, skills, scheduling, triggers, integrations, voice | Everything in V1, plus ↓ | Everything in V2, plus ↓ |
| Acts as you | — | Email, meetings and phone in your voice; live meeting attendance; cloud computer use; knowledge mindmaps; CodeAct | ↑ |
| Multi-user and governance | Single-user login | Multi-user login and invites, admin console, per-user usage and cost, encrypted credential vault, skill governance, Agent API and Agent Mesh, service verticals | ↑ |
| Works with you | — | — | Your own computer; Task Board and jobs that keep working over days; teach a task by doing it; the Digital Twin — standing to disagree, and autonomy that is only ever earned |
The shift each time is about who does the finishing. V1 does the work you hand it. V2 does the work as you, to other people. V3 holds work you haven't handed it yet — on your machine, over days, and with the standing to tell you when you're drifting.
Not sure what's available on your instance? Just ask the agent — "what can you do?" — and it answers from its own current capabilities. See Ask the agent about itself.
What it can do
Breaks a goal into steps and runs them — research, code, files, messages, scheduling — correcting itself along the way, and stops only when the result is actually there.
Joins a Teams, Zoom or Google Meet link, listens, responds in your cloned voice, and sends you a transcript with action items.
Reads inbound email, Slack DMs and Telegram messages. Drafts replies in your writing style. Waits for your approval before sending.
Works to your schedule without being asked — morning briefings, market checks, goal reviews — and pushes results to your phone.
Your files, your scripts, your installed apps. Folder-by-folder permission, a banner on the machine while it works, and it hands the mouse back the moment you touch it.
A cloud browser that keeps your sign-ins. Show it a task once and it becomes a skill you can run by name — or hand to a job.
A Task Board card becomes a job: an agent that wakes on a schedule, an email, a calendar event or a webhook, and remembers everything it already tried.
Saves what worked as a reusable skill. The next time the same job comes round it runs the known solution instead of working it out again.
GitHub, Notion, Jira, Gmail, Slack, Stripe, HubSpot, AWS S3, Twitter and 33 more. Connect one and its tools are immediately available.
Your history, the facts it has learned about you, everyone you deal with across every channel, and your own documents. Context persists across weeks.
Spawn up to 5 sub-agents for complex tasks — research in parallel, coordinate results, deliver a single unified output.
Agent Mesh: discover any A2A-compatible agent on the internet, delegate tasks with cryptographically scoped authority, classify behavior in real time, and maintain an immutable audit trail — with canary tripwires and automatic HOSTILE escalation.
Who uses it
The problem: You're double-booked. Every missed meeting, dropped follow-up, or late competitive check has a direct cost.
What changes with AI Partner:
- Can't attend both calls? Partner joins the second one, drives toward the goal you defined in
GOALS.md, sends you an outcome summary. - Follow-ups dying in someone's head? Action items are extracted from every meeting and reviewed via Telegram before anything is sent.
- No time for competitive research? Heartbeat runs every morning, delivers signals to your phone before your first call.
- Goals scattered across Notion and your head?
GOALS.mdis the single source — every meeting and proactive task aligns to it automatically.
Setup: 20 minutes. Fill in USER.md, write GOALS.md, set HEARTBEAT.md.
The three things that make it different
A goal isn't done when the agent stops talking — it's done when the thing you asked for exists and has been checked. A run that produced a summary instead of the deck you asked for counts as incomplete.
Anything touching your real accounts goes through your own authority policy first: do it, draft it and ask you, or never. You set the line; it doesn't move on its own.
What you said last month, who owes you what, the decision you made in March — all still there. Memory is the feature, not the context window.