Technical Report · July 2026

The Room Protocol

Shared-State Coordination for Multi-Agent Software Development

Shreyas Padmakiran · roomd.sh

Abstract

Large language model coding agents are effective in isolation but have no native way to coordinate when several of them work on one software project. The common workarounds, a human relaying state, agents exchanging messages, or agents sharing only a repository, all treat coordination as message-passing. This work argues the better primitive is shared state: a small set of structured, persistent, queryable objects that agents read and write instead of messaging one another.

The Room Protocol organizes coordination around a single abstraction, the room, which namespaces five state primitives (a plan, a typed context store, an append-only event log, presence, and locks) and is the unit of ownership and access control. It is exposed over the Model Context Protocol, so any MCP-capable agent can participate without a custom client. The reference implementation, roomd, is a fully stateless server backed entirely by a single Redis: the properties that make coordination durable are the ones that make the system multi-tenant and deployable.

roomd vs the Room Protocol

One idea, one thing that runs it. The same split as HTTP versus a web server.

the Room Protocol
The design: the room abstraction, five state primitives, exposed over MCP. A spec, not a program.
roomd
The reference implementation: the stateless server you run and point agents at.
roomd-web
The dashboard humans log into to watch and manage rooms.
room
The core primitive: one workspace, named by a roomId, that all coordination state lives under.

Contributions

  1. 01The Room Protocol: one abstraction, the room, with a primitive set argued sufficient to coordinate independent coding agents.
  2. 02A typed-context model: context entries carry a type and a per-type schema, so a consumer relies on shape rather than parsing prose.
  3. 03A stateless, single-store reference architecture: a fresh MCP server per request over one Redis, durable and restartable.
  4. 04A layered multi-tenancy and access model: three secret types, first-touch room ownership, and rate limiting that fails open.
  5. 05A concurrency mechanism: a distributed plan lock, plus per-agent cursors that give each agent an exactly-once view of events.
  6. 06An evaluation methodology for coordination overhead, handoff correctness, and concurrency safety.

Read the full report

The complete paper covers the protocol operations, the typed-context model, the stateless architecture, concurrency and access control, an evaluation methodology, and related work.