Ankole

The open source Agent Harness with a Company Brain

Website · Documentation · Quick start · Discussions

English · 简体中文 · 日本語 · 한국어

Ankole runs AI Agents that do ongoing work for a company, from a one-person business to an enterprise. People reach them in the chat tools they already use, such as Slack, Microsoft Teams, Lark, DingTalk, and WeCom.

A model with a prompt and a few tools starts every task from zero. It does not know what your team decided last week, it acts with whatever access it holds, and its work ends when its process dies. Ankole is the harness that fills these gaps. Its Agents share a Company Brain that remembers what teams decide. They work within the access you grant, and they recover long tasks after crashes and restarts. Each Agent has its own Linux computer in a container, and you run the whole system on infrastructure you control.

Ankole is an open source alternative to Claude Tag, the Slack agent that Anthropic hosts. Ankole is not tied to Slack or to Claude models, and it keeps company context on your servers. Model and search calls go only to the providers you configure, which can include a model that you host yourself.

#research · group chat

Ravi: for cyclical stocks, use data from a full cycle; three years is too short
Mia: agreed, 7 years minimum

Three days later

Sam: @Ankole run a valuation on the metals stocks
Ankole: On it. Using seven years of data for the cyclical stocks, as this channel agreed three days ago.
Attached: metals-valuation.pdf (14 pages)

Nobody asked Ankole to remember that rule. Company Brain learned it from a chat that Ankole belongs to, kept its source and date, and applied it to the next request that needed it.

Why Ankole

A model supplies the reasoning. The harness around the model decides what the model sees and what it may do. The harness also decides how work continues after a failure and how results reach the next decision. Ankole puts these decisions into the system instead of into each prompt:

• Agents remember what the company decided. Company Brain learns from conversations, registered files and URLs, and notes that Agents write. It stores each piece of knowledge as a claim with its source, time, confidence, audience, and holder: the person or Agent whose view it is. New evidence updates the current view and keeps the history, and conflicting claims stay visible for review.
• Each Agent works within the access you grant. People and Agents each have an identity and permissions, and Agents share no computer, home directory, or credential. Before the model reads company knowledge, Company Brain removes anything that the asker has no permission to see; in a group chat, it can apply the permissions of every member. PostgreSQL records what each Agent was asked, the model calls it made, and what it delivered. Token quotas cap the model spend of each Agent. Offboarding a person blocks their sign-in and stops the tasks and schedules they own.
• Work starts when something happens. A message, a schedule, a webhook, or another external event wakes the Agent that owns the work. An Agent can also set a reminder to check back later.
• Long work survives failures. A task can run for hours or days and accept new input while it runs. The control plane runs on Elixir/OTP, which confines a crash to the conversation where it happened while other work continues. PostgreSQL keeps the inputs, model history, and tool results that Ankole uses to continue the work.
• Each Agent has its own computer. A Linux container gives the Agent a browser, a terminal, files, Codex, MCP (Model Context Protocol) servers, and Skills: approved procedures for one kind of work. Its home directory survives restarts and upgrades.
• Corrections and outcomes feed later work. Deep Research gathers evidence in several separate contexts, tests competing hypotheses, and returns a cited report. After repeated corrections, an Agent can propose a Skill lesson: a short rule added to a Skill, which takes effect when a person approves it. When a forecast recorded in Company Brain comes due, Company Brain compares it with what actually happened.

Ankole is a complete system for company work, not a library. It brings its own control plane, database, and Agent computers. For an agent loop inside your own application, or a personal assistant for your private life, a smaller tool will fit better.

What people ask Ankole to do

These requests arrive in a group chat or a direct message:

Works with

Quick start

Docker Compose runs a complete deployment on one Linux, macOS, or Windows host. The stack contains PostgreSQL, the control plane, and Caddy as the HTTPS server. You need:

• Docker that can run Linux amd64 or arm64 containers, with memory and disk for one container per running Agent.
• A host where Ankole may have root access. The control plane creates the Agent containers through the Docker socket, and access to that socket equals root access on the host.
• A DNS name for the host and open ports 80 and 443. For a trial on your own machine, use the name ankole.localhost instead.
• An API key for a model provider, and permission to create a bot on your chat platform. Ankole connects out to Slack, Lark, DingTalk, WeCom, Discord, Telegram, and email, so they work in a local trial. Microsoft Teams, LINE, and WhatsApp send events to the host and need a public HTTPS address.

Get the deployment files:

git clone https://github.com/AgentBull/ankole.git
cd ankole/tools/deploy/docker-compose
cp .env.example .env
chmod 600 .env

Set these values in .env:

Start the stack and read the one-time activation code:

docker compose pull
docker compose up -d
docker compose logs control-plane | grep "SETUP ACTIVATION CODE"

The code appears when the control plane has started. If the last command prints nothing, wait a moment and run it again.

With ankole.localhost, Caddy signs the certificate with its own local authority, and the setup page needs trusted HTTPS. Copy the root certificate of that authority, then trust it with the certificate tool of your operating system:

docker compose cp caddy:/data/caddy/pki/authorities/local/root.crt ./ankole-local-ca.crt

Open https://<ANKOLE_HOST>/setup and enter the code. Choose the plugins that you need, such as the one for your chat platform. Set up how administrators sign in: through an identity provider such as Slack or Microsoft Entra ID, or with a 6-digit email code sent through your SMTP service. Then, in the Console, the web interface for administrators, add a model provider, create an Agent, and connect a chat platform to it. Keep provider API keys in the Console, not in .env. Send the Agent a message in that chat; its repl