A living, agent-maintained understanding layer for OpenAB — the open Agent Client Protocol broker.
This repo is auto-updated by an agent workflow that watches beta releases on openab main, diffs what changed, and rewrites affected sections. Every file traces back to a source commit SHA in .sync-state.
If you're new → start at What is OpenAB
| Layer | What it answers |
|---|---|
| 00 · What is OpenAB | One-page pitch. The problem, the solution, what it is not. |
| 01 · Core Concepts | The 10 ideas you must internalize before anything clicks. Includes the OAB MCP Facade, Agent Control Plane (Phase 1 released in beta.5), reactions mapping, and Sandboxes & Computer Grants (ecosystem: openab-pty + instance-mcp). |
| 02 · Mental Models | How the pieces fit — data flows, topology, sequences. |
| 03 · Use Cases | "I want to do X" → here's how. Includes driving an agent from an ACP client, browser control from chat, PR contribution lifecycle, and the Jelli governed community bot setup. |
| 04 · Decision Trees | Should I use X or Y? Structured branching choices. |
| 05 · Change Digest | Plain-English summary of recent openab changes. |
openab/beta push or daily schedule
↓
GitHub Action: compute diff since last sync
↓
Claude agent: reads diff + current map
↓
Rewrites affected sections + updates LATEST.md
↓
Opens PR
↓
Multi-agent review via openab b2b
↓
Auto-merge on consensus
The agents doing the review run through openab itself — the same system being documented. See the repo's agent sync workflow.
- Repo:
openabdev/openab - Branch:
main(beta release tags, currentlyv0.10.0-beta.5) - Sibling repos tracked for the sandbox/computer-grants concept:
openabdev/openab-pty,openabdev/instance-mcp - Last synced SHA: see
.sync-state
The map is organized by concept, not by repo — there is no per-repo file. Use this index to find where a given source repo's behavior is documented:
| Source repo | Primary coverage | Also referenced in |
|---|---|---|
openabdev/openab |
Everything under layers 00–04 | Change Digest |
openabdev/openab-pty (sandbox runtime: locked-down container shells over a WireGuard tailnet, oab-toolchain image) |
Sandboxes & Computer Grants — Piece 1 | Change Digest status check; Trust Model "Beyond Identity" |
openabdev/instance-mcp (computer-side executor: reverse attach, tool-profile grants, switchboard mode) |
Sandboxes & Computer Grants — Pieces 2–4 | Change Digest status check; Trust Model "Beyond Identity" |
This map is agent-maintained but human-correctable. If a section is wrong or missing:
- Open an issue describing the gap
- Or open a PR — the next agent sync will reconcile your changes against the source
Priority areas that benefit most from human input: decision trees, use-case narratives, and "gotchas" that aren't visible from the code alone.
