A personal AI companion you host yourself.
It remembers you, shows its work, and never acts without your say.
Overview · How it works · Using it · Running it · Status
- Talk to the models you choose. Local models through Ollama, hosted ones through OpenRouter, OpenAI or Anthropic, or sign in with a ChatGPT or Claude plan. Answering, thinking and reviewing each get their own model, or let Auto pick.
- Remembers, with sources. Every memory keeps where it came from. Correct or forget any of them from the Memory screen.
- Acts only as you allow. Every action goes through ToolGate. Your settings decide, per chat, agent and tool, whether it runs, the AI reviewer decides, or you are asked. When the reviewer decides, anything that reaches a real person or can't be undone still comes to you.
- One Inbox. Approvals, Conker's own requests (a cron job, a tool, a memory) and its ideas wait in one place, one card each: do it, don't, or change it.
- Works while you're away. Cron jobs are your own instructions, run by the AI with tools on a schedule, an event or a webhook, within the budgets you set.
- A terminal for agents, not for your server. Each command runs in a fresh locked-down container with no network; packages arrive only through a read-only proxy.
- Everything is a setting. 102 settings that take effect (221 more planned), one pattern,
searchable. The same settings work from the terminal with
conker config.
 Home. What needs you, what's coming up, how much of each plan is left. |
 Inbox. One card per decision, with the buttons that fit it. |
 Memory. What Conker knows, and the chat each fact came from. |
 Settings. Every setting by topic, in one pattern. |
Screenshots show the built-in preview with sample data.
See it in five minutes. The interface runs on its own with sample data. Needs Git and Node.js 22.12+.
git clone https://github.com/Conker-AI/conker.git
cd conker/dashboard
npm ci
npm run dev -- --host localhost --port 5173 --strictPortOpen http://localhost:5173. Answers, memories and approvals are simulated in the browser. CI runs
these lines as written on every dashboard change and fails if the page is not up within five
minutes (scripts/check_quickstart.py).
Install on a server. Any Linux machine with Docker (WSL2 works too). 16 GB of RAM is comfortable for local models.
git clone https://github.com/Conker-AI/conker.git
cd conker
./install.shThe installer checks the machine, generates every secret, pulls the pinned images and prints one
address. It then offers ./conker setup, a guided first run that asks one plain question at a
time. Day to day:
./conker status # is everything working, and what isn't
./conker update # a pinned release, verified, with rollback
./conker config list # every setting, from the terminal
./conker backup # a verified snapshotOther ways to run it, including a native development stack on Windows and the source-built Ubuntu layout: 4 · Running Conker.
Conker is early. The current source has 40 of 41 tracked capability families active; the owner's own terminal is still deferred. Local models on a CPU work but are slow; a hosted or larger model makes tool use and drafts much more reliable. What works today
Small services with one job each, so every boundary is a process boundary. Each has its own repository and runs on its own.
flowchart LR
You([You]) --> GW[Gateway<br/>login, approvals]
GW --> Pi[Pi<br/>conversations]
Pi --> MG[MemoryGate<br/>remembers]
Pi --> TG[ToolGate<br/>acts, as allowed]
Pi --> SG[SystemGate<br/>observes]
Pi --> M[Models<br/>local or hosted]
MG --> EM[Embeddings]
TG --> SB[Sandbox<br/>agent terminal]
| Repository | Job | Release |
|---|---|---|
| pi | Conversations, model roles, the gateway and a record of every turn |  |
| toolgate | The only way to act: typed tools, exact approvals, spending limits, the sandbox |  |
| memorygate | Memory with sources: evidence in, bounded context out |  |
| systemgate | A read-only view of the machine |  |
| embeddings | Text to vectors, and nothing else |  |
This repository holds the app, the installer, deployment and the docs. versions.env
pins the exact release of each service. More in 2 · How it works.
| 1 · Overview | What Conker is, why, and the words it uses |
| 2 · How it works | Services, one message, one approval, in diagrams |
| 3 · Using Conker | Each screen and common tasks |
| 4 · Running Conker | Install, update, back up, recover |
| Looking after your server | Update, back up, restore and roll back, in plain steps |
| 5 · Developing | Repositories, checks, rules, decisions |
| Status | What works today, and what doesn't yet |
The frontend builds on shadcnstore's dashboard template (notice), React, Vite, Tailwind CSS and Radix.
AGPL-3.0, version 3 only. Copyright (C) 2026 alexeybe1kin. Releases published before 24 October 2026 were under MIT and stay under MIT. Found a vulnerability? Please report it privately: security policy.