v1.1.0 · Apache-2.0 · in daily use since November 2026
agent-kernel
The agent in one container.
A Claude Code-based 24/7 personal agent on Telegram, using only official components: the Claude Code harness, the official Telegram channel plugin and a managed policy file. Memory, scheduled jobs, groups and fellow agents, with less than 300 lines of Bash and no custom agent code.
One message, start to finish
You, in your team's group
Read yesterday's sales and tell each rep what is still pending.
Official plugin
claude
Reply tool
Your team
Ana: 3 follow-ups due today. Luis: 2 orders waiting on stock. Full list attached.
An example, with made-up names. agent-kernel reads Excel, Word, PowerPoint and PDF files with no setup, makes new ones and sends them back, and answers everyone in a group you add it to. What it can do · Groups
Three ways to run it
Railway
One click from the pixan-agent template. You add one volume at
https://railway.com/deploy/pixan-agentRailway guide/dataand three variables: a Claude credential, the bot token and your Telegram user id.Docker
Create a bot with @BotFather, get a Claude credential, then build and start with the four commands below.
Step-by-step guideYour own Linux server
Any amd64 Linux server with Docker, and no incoming port. Install, updates and backups are in the guide.
Linux guide
git clone https://github.com/pixan-ai/agent-kernel cd agent-kernel cp .env.example .env # fill in one credential and the two Telegram values docker compose up -d --build
Use a new bot token, and expect the first build to take several minutes.
Published images
Every release publishes the base image: the operating system, tools, policy and scripts, with no Claude Code in it. Your own build adds Claude Code on top, so the binary is never redistributed.
docker.io/pixan/agent-kernel:1.1.0 ghcr.io/pixan-ai/agent-kernel:1.1.0
Images are signed with cosign and carry an SBOM and SLSA provenance. How to verify them · Docker Hub · What is published and why
What it protects, and what it does not
Defense in depth, each layer an official mechanism: a sender allowlist, a root-owned managed policy, the permission classifier, hooks that tell the owner about every change to the agent's tools, and a non-root container.
It has not had an independent security audit. Treat the shell as unconstrained, expect a soft floor only on Railway, and note that there is no egress control by default. Read the threat model before trusting it.
Documentation
- Getting startedFrom nothing to an agent that answers, step by step.
- ArchitectureWhat runs in the container, and why there is no supervisor.
- Threat modelWhat it protects and what it does not.
- FindingsThe evidence behind each claim.
- LearningsWhat earlier fleets and this project taught us.
- MemoryThe daily dream: consolidate memory, then start clean.
- Scheduled jobsPrompts that run on a schedule and deliver the answer.
- GroupsTaking part in Telegram groups.
- Fellow agentsWhat agents may do for each other over SSH.
- SleepSleep when idle, wake on the next message.
- BackupNightly encrypted backup, and how to restore.
- ReleasingHow an image is built, signed and published.