What Brainstack is
Brainstack gives agent harnesses a shared place to read context and write useful evidence without turning every note into trusted memory immediately. The source of truth is a git-backed shared brain. braind serves search and proposal APIs, while brainctl handles install, enrollment, diagnostics, local capabilities, hooks, skills, and fleet operations.
Brainstack is not another chat harness. Codex, Claude, Cursor, Telegram, scripts, and workers can all use it from their existing surfaces. If telemux is configured, Telegram can be an operator surface too: messages like install voice on worker-a and commands like /updates, /curator_run, or /proposals pending route to the same Brainstack operations as the CLI.
The basic loop
- A harness asks Brainstack for local project context.
- The harness does work in the normal tool or machine.
- Hooks, commands, or operators post raw evidence back to Brainstack.
- The curator turns evidence into proposed wiki or memory edits.
- An operator accepts useful proposals and rejects stale or vague ones.
Good first paths
- New Mac client: start with Install Brainstack.
- Evaluating locally: run the Local proof before enrolling a real machine.
- Existing operator: review Proposals and curation.
- Local power-up: install Capabilities.
- Harness setup: check Skills and harness context.
- Mac operator surface: use the Brainstack Menu app to watch health, proposals, Telegram routing, and fleet updates.
- Pilot review: read the Trust model before exposing services or inviting more machines.
Trust model in one paragraph
Brainstack is designed for trusted private networks. In the default tailnet posture, network reachability is the read boundary, so Tailscale ACLs, firewall policy, and SSH host pinning matter. Do not expose braind directly to the public internet. Clients should write imports and proposals, not direct wiki edits. Admin tokens and Telegram bot secrets belong on the control host or operator machines, not on every client.