Skip to content

For AI Agents: the documentation index is at https://ankole.agentbull.com/en-US/llms.txt.

Ankole
Open source · Self-hosted on your infrastructure · From solo teams to enterprises

The Agent Harnesswith a Company Brain

Ankole unites a Pod digital headquarters—collaborative Docs, sharded Bases, React Applets, and Pod Chat—with Personal Assistants and specialist AI Colleagues. Powered by Meta Ankole's Company Brain, Skills, and unified Connectors, they work across Pod and Slack, Teams, Lark, DingTalk, or WeCom.

research9 members
  1. Ravi14:02

    for cyclical stocks, use data from a full cycle; three years is too short

  2. Mia14:03

    agreed, 7 years minimum

Ankole picked up a team rule without anyone explicitly teaching it

Three days later

  1. Sam09:41

    @Ankole run a valuation on the metals stocks

  2. Ankoleagent09:41

    On it. Using seven years of data for the cyclical stocks, as this channel agreed three days ago.

    metals-valuation.pdf14 pages · checked
  • Long-running tasks survive crashes and restarts
  • Pod workspace + external IMs, with no model lock-in
  • Personal Assistants & AI Colleagues with strict runtime permissions
  • Apache-2.0 open source

01Use cases

What teams ask Ankole to handle

Work starts in Pod Docs, Bases, Applets, and Pod Chat—or directly from Slack, Teams, Lark, DingTalk, and WeCom.

  1. Send a weekly digest

    Creates a recurring Friday schedule, reads posts from the nine X accounts you specified, groups them by topic, and compiles a PDF.

    A formatted PDF posted to Pod Chat or your channel every Friday at 17:00

  2. Fix an issue

    Delegates to an engineering AI Colleague running Codex in a background job via unified GitHub and Linear Connectors.

    A verified pull request linked back into the conversation

  3. Sync trackers and recap the week

    Runs published zero-LLM Applet Actions to sync Lark, GitHub, and Linear updates into a sharded Base, then summarizes open items from authorized chats.

    Live Base rows, an interactive Pod Applet view, and a weekly checklist

  4. Build slides or translate documents

    Turns rough notes into a presentation deck or translates a PDF while preserving layout.

    A ready-to-share PowerPoint deck or translated PDF

  5. Verify a market rumor

    Launches a Deep Research job to gather and weigh primary evidence on both sides of the claim.

    A cited brief laying out the supporting and contradicting evidence

  6. Replicate an index

    Fetches the underlying series and recomputes the index from its published methodology in a Jupyter notebook.

    A comparison chart plotting the replicated series against official figures

Fluent text is cheap; every request should end in verifiable work.

02Company Brain

Nobody told Ankole to memorize the seven-year rule

Company Brain picked up the rule from a channel Ankole belonged to, recorded its source and date, and applied it automatically the next time a valuation came up. Every authorized Agent works from the same up-to-date company knowledge.

Company Brain · Research rules4 rules

  1. 07-14Cyclical stocks use percentiles from seven years of dataAgreed in a conversation between two colleagues
  2. 07-19Lead with the conclusion and include the risksAfter a draft came back
  3. 07-22State when data is missing and use only observed valuesWritten into the skill after an incident
  4. 07-25Value the HK listing separately from the A share listingCorrected once

Rule lifecycle in Company Brain

Plain vector search retrieves raw message snippets by embedding similarity—often returning an outdated rule right alongside the newer rule that replaced it.

Company Brain instead stores each piece of knowledge as a structured claim with its source, timestamp, confidence, disclosure scope, and holder (the person or Agent behind that view). When new evidence arrives, Company Brain updates the active view while keeping historical claims and surfacing unresolved contradictions for review.

Every new request starts from what your company knows today. Repeated feedback sharpens operational rules, and every forecast stays linked to the real-world outcome that tests it.

We run Company Brain on Ankole's own codebase

An Ankole Agent helps maintain this repository using the exact same Company Brain, tools, and governance rules that ship to users. Every review correction and test outcome refines the context and procedures used in later maintenance.

03Harness

A bare model with a prompt and a few tools starts every task from scratch

Left on its own, a model forgets what your team decided last week, uses whatever credentials happen to be nearby, and loses all progress if its process dies. Ankole's harness closes those gaps.

  1. 01

    Agents remember what your team decided

    Meta Ankole's Company Brain learns from Pod Docs, sharded Bases, team chats, and registered URLs. Every ontology page and claim retains its source, date, and audience scope—updating the current view as new evidence arrives while preserving full history.

  2. 02

    Corrections and real outcomes improve future work

    From real Job trajectories, Dreaming automatically activates 7-day leased Skill cautions to prevent repeat tool or process mistakes, while humans curate permanent cautions and shared Skills in Meta Ankole. Company Brain also grades forecasts against actual outcomes.

  3. 03

    Work wakes up on real events

    Chat messages, cron schedules, webhooks, and Connector triggers wake Agents automatically—or execute published zero-LLM Applet Actions on Function Runtime for deterministic checks and syncs.

  4. 04

    Long-running work survives crashes

    Tasks can run for hours or days while accepting follow-up steering along the way. If a Session process fails, Ankole resumes it from the inputs, model tape, and tool receipts committed to PostgreSQL.

  5. 05

    Every Agent stays within the access you grant

    Every member gets a Personal Assistant scoped to their collaboration permissions, while specialist AI Colleagues use their own Principal grants, Peer A2A delegation, and Action Card approvals. Unified Connectors keep credentials in the control plane, and offboarding halts owned work immediately.

04Capabilities

Built for ongoing company operations

Ankole is a complete digital headquarters and Agent Harness—shipping Pod workspaces, Meta Ankole, unified Connectors, Function Runtime, and dedicated Agent computers.

Work keeps running asynchronously

Deep research jobs, scheduled checks, and zero-LLM Applet Actions keep running long after you close the chat—and resume cleanly from PostgreSQL after a restart.

Shared knowledge across Pod and chat

Meta Ankole distills working rules, ontology pages, and ruled-out options from Pod Docs, sharded Bases, Pod Chat, external IMs, and registered URLs so Personal Assistants and AI Colleagues share current context.

Permission-aware knowledge disclosure

Before any company knowledge enters a model prompt, Company Brain intersects access permissions—enforcing participant audience rules across Pod pages, group chats, and Peer A2A exchanges.

Deep Research tests competing hypotheses

Deep Research gathers evidence across isolated sub-contexts, weighs competing explanations against each other, and delivers a cited report.

Every Agent gets its own dedicated computer

Each Agent runs in its own Linux container with /drive FUSE access to Unified Drive, a headless browser, terminal, Codex runtime, Skills, and control-plane Connectors—backed by a persistent /workspace home.

Meta Ankole Skills and leased lessons

Meta Ankole manages shared Skills plus AGENTS.md and DESIGN.md contracts. Dreaming automatically activates 7-day leased cautions from real Job trajectories, while humans curate permanent cautions and Skills.

Peer A2A, Workflows, and Pod Applets

Personal Assistants delegate to specialist AI Colleagues via isolated Peer A2A sessions, Workflows fan tasks across independent contexts, and React + TypeScript Applets run zero-LLM Actions on Function Runtime.

Full audit trail and per-Agent token quotas

PostgreSQL records accepted conversation inputs, model invocations, and delivery outcomes. Model Profile defaults with per-Agent overrides and per-Agent token quotas govern model spend.

05How it works

Built so enterprise work survives process failures

Virtual Actor identities, OTP fault isolation, live Worker channels, disposable compute, and PostgreSQL transactions keep long-running work recoverable.

  1. 01

    Sessions keep their identity across restarts

    Every active Session runs as a Virtual Actor with a stable address, durable state, mailbox, lifecycle, and recovery from committed progress.

  2. 02

    Supervised recovery for each Session

    Elixir/OTP supervises each Session in its own execution process, isolates failures in that process, and restores committed progress from PostgreSQL after a restart.

  3. 03

    Live control over running Agents

    A dedicated Phoenix Channel per Worker multiplexes environment tool calls, Codex Job control, checkpoints, and backpressure while the control plane drives the main conversation loop.

  4. 04

    Dedicated, replaceable Agent computers

    Agent Computer runs shell commands, Codex Jobs, headless browsers, /drive FUSE workspace mounts, and persistent /workspace files, while deterministic Applet Actions run on Function Runtime. Worker process state can be rebuilt at any time because PostgreSQL owns committed domain facts.

  5. 05

    PostgreSQL owns durable completion state

    Messages, turns, schedules, authorization decisions, and delivery intents commit transactionally to PostgreSQL. Live streams show progress; PostgreSQL defines what actually happened.

06Run it

Run the full Harness

Launch on a single Linux, macOS, or Windows host with Docker Compose, deploy to Kubernetes in production, or run straight from source.

Run this Compose stack on any Linux, macOS, or Windows machine capable of running Linux Docker containers. It brings up PostgreSQL, Function Runtime, the control plane, and Caddy HTTPS, and launches one dedicated Worker container per Agent.

bash
git clone https://github.com/AgentBull/ankole.gitcd ankole/tools/deploy/docker-composecp .env.example .envchmod 600 .envFill the required values in .env, then:docker compose pulldocker compose up -d

Give your company a shared Brain

Self-host Ankole on infrastructure you control. Your company's context is stored on your servers, and model or search requests go only to the providers you configure.