DocumentationChangelog
Getting Started

Introduction

Talon Control is the operating system for running a software company — customers, licences, money, mail, publishing and an AI agent fleet in one self-hosted appliance. Single container, single Postgres, single operator.

What Control is

Control began as a customer registry and licence-key issuer, and that is still its spine: it is the book of record for who your customers are and what they hold. Around that spine it grew the rest of running a business — the work, the money, the mail, the memory, the statutory registers, and the AI fleet that does a share of it.

It is deliberately an appliance, not a platform: one container, one Postgres database, one operator. You run it on your own infrastructure, your data never leaves it, and your licence verifies offline. We run Talon Labs itself from the same image you install.

The four parts

Book of record
Customers, signed licence keys (minted, revealed under audit, revoked), the registry of hosted deployments you operate, and an append-only event log behind all of it.
The business
A full UK double-entry ledger, CRM with contacts → deals → pipeline, Gmail-mirrored mail, outbound sequences behind compliance guardrails, content publishing that commits into your site repo, cookieless traffic analytics, scheduling with public booking links, and the employee roster with per-section access control.
The fleet
Named AI agents with curated tool grants, a Coordinator that plans each day, built-in automations on a minute tick, a voice-first stand-up, dev sessions that drive coding agents against real git checkouts, and the Company Brain — a provenance-weighted note corpus retrieved into every run.
Governance
The registers a real company has to keep and usually does not: records of processing, vendor DPAs, the breach log, statutory registers, the IP and insurance registers, right-to-work checks, contracts, access-review cycles and a findings register — with a compliance dashboard reading across all of them.

Because all four share one database, a deal, its invoice, its payment, the customer's email thread and the processing record covering it are the same record seen from different sections — there is no sync, and there is nothing to fall out of sync.

Agents propose, humans approve

The whole fleet is built around one invariant: agents propose, humans approve. Drafted replies, plans, cards and proposals arrive ready to review; sending, deploying and spending remain human acts. Outbound tools are physically absent from autonomous runs rather than merely discouraged — there is no configuration in which an agent emails your customer on its own.

The same line as Talon Studio
Studio's detection agents hold the identical invariant — every agent proposal is a reviewable change waiting for a person. It is the company's public position, not a config value.

What you actually run

Control ships as a published container image. You run it on a small VM you control — a €5/month cloud box, a Railway or Render service, a machine in your own rack — next to a Postgres database, behind your own reverse proxy. There is no hosted tier to sign into, no control plane phoning home, and nothing to install from source.

Two vCPU and 4 GB of RAM is a comfortable starting point, and one instance is the supported shape — see Deployment & security for why the replica count is a correctness question rather than a capacity one.

Where to next?