AgentRoom (ar) is the Chatto-native merge desk for agent-authored pull requests.
AI coding agents can now create branches and pull requests faster than humans can supervise them. The bottleneck is no longer writing code. It is knowing which agent PR is ready, blocked, risky, or waiting on human judgment.
AgentRoom uses Chatto as the live human/agent collaboration surface and GitHub as the source of truth. Each repo gets a Chatto room. Each agent-authored PR gets a thread. AgentRoom posts intent, plans, branch links, CI state, blockers, evidence summaries, and approval requests through a dedicated Chatto connector account. Humans approve, reject, pause, or redirect from the room. AgentRoom records the decisions, applies readiness policy, tracks GitHub state, and posts evidence packets back to PRs.
The technical bet is simple: combine Chatto's self-hosted collaboration UX and NATS-backed realtime foundation with GitHub's code-review and branch-protection model.
Chatto is pre-1.0 and does not currently appear to expose a dedicated bot/webhook integration API. AgentRoom should treat Chatto integration as an adapter boundary, not a hard-coded internal dependency.
Initial dependency model:
- Bring your own Chatto instance.
- Configure AgentRoom with a Chatto URL and a dedicated Chatto user credential.
- Post messages through Chatto's public ConnectRPC
MessageService. - Detect human decisions through Chatto's realtime WebSocket where possible, with timeline polling as a fallback.
- Keep GitHub as the source of truth for PRs, checks, reviews, and merge.
See docs/chatto-integration.md for the current integration assumptions and risks.
Tagline: Every repo gets a room. Every agent PR gets a decision.
AgentRoom is in early open-source planning. This repository currently contains the product brief, architecture direction, safety model, and pilot workflow templates. The first executable MVP will focus on one GitHub repo connected to one Chatto room, with every agent-authored PR classified as needs review, blocked, risky, or ready.
- GitHub maintainers running multiple coding-agent tasks per week
- Tech leads and staff engineers reviewing agent-generated PRs
- Platform and DevEx teams building internal agent workflows
- Technical founders who want more agent throughput without giving up control
- A Chatto-native room for each GitHub repo
- A PR thread for each agent-authored pull request
- A merge-confidence state machine for agent work
- A human approval and policy checkpoint
- A PR evidence packet generator
- An append-only audit trail across agents, humans, CI, Chatto, and GitHub
- Not a foundation model
- Not an IDE
- Not Jira
- Not a general team chat product
- Not an autonomous merge bot
- Not a replacement for GitHub pull requests
The first version should prove one narrow outcome: connect one GitHub repo to one Chatto room and make agent-authored PRs easier to decide on.
Planned MVP components:
- GitHub App
- Experimental Chatto connector
- Repo room binding
- PR thread creation
- Readiness state machine
- Human approval prompts in Chatto
- PR evidence packet
- Append-only event log
- Generic agent event webhook or one runner adapter
Read the getting-started guide:
open docs/getting-started.mdIf you want to pilot the workflow before the executable MVP exists, use the templates:
| Path | Purpose |
|---|---|
docs/getting-started.md |
Contributor and pilot workflow guide |
docs/product-brief.md |
Product pitch, target user, wedge, and MVP |
docs/architecture.md |
Initial system architecture and data flow |
docs/agent-adapter-contract.md |
Generic agent event contract and first adapter choice |
docs/chatto-integration.md |
Chatto dependency model, connector assumptions, and integration risks |
docs/chatto-message-format.md |
Chatto thread messages, approval commands, parsing rules, and ambiguity handling |
docs/github-app-design.md |
GitHub App permissions, evidence packets, readiness checks, and fork behavior |
docs/github-app-setup.md |
Manual GitHub App creation, installation, private key, webhook, and config setup |
docs/evidence-packet-format.md |
Sticky PR comment, check run body, packet versioning, and partial-data rendering |
docs/readiness-evaluator.md |
Deterministic readiness decision tree and conflict resolution |
docs/runtime-architecture.md |
MVP process model, SQLite state, connectors, workers, and deployment |
docs/state-model.md |
SQLite event log, latest-value state, evaluator read path, and idempotency |
docs/configuration.md |
TOML config schema, required fields, secret handling, and init behavior |
docs/test-strategy.md |
Unit, connector, renderer, fixture E2E, idempotency, and live-test strategy |
docs/security.md |
Trust, safety, and approval model |
docs/roadmap.md |
Initial development milestones |
docs/superpowers/specs/2026-07-09-cross-harness-agent-review-room-design.md |
Future cross-harness agent review room extension |
docs/templates/ |
Manual pilot templates |
MIT