01Agent toolingOpen source

Houseparty

Local-first multi-agent orchestrator for Codex and Claude Code, with persistent terminals, shared memory, and remote supervision.

← Back to projects

Houseparty runs coding agents as persistent, supervised sessions instead of throwaway subprocesses, with PTY-backed terminals, shared memory, and resumable runs.

System proof

Architecture proof

Houseparty is not just a prompt wrapper around coding agents. The repo already shows the real system boundary: a Rust daemon supervising PTY-backed Claude and Codex sessions, a Next.js dashboard over REST and WebSockets, project-scoped memory and MCP translation on disk, and Telegram as a remote-control surface.

Core runtime

Rust daemon + Next.js dashboard

Agent model

Persistent PTY sessions, not one-shot subprocesses

State on disk

.houseparty projects, agents, MCP config, message log

Control surfaces

Browser dashboard, gamemode, and Telegram remote control

Houseparty

Problem

Most agent frameworks treat the CLI as a one-shot subprocess. They run a command, capture stdout, and disappear. That makes it difficult to intervene mid-flight, preserve context, or supervise multiple long-running agents in a coherent way.

Houseparty was built around a different assumption: coding agents should be treated as durable working sessions, not stateless invocations.

Houseparty

What it does

Houseparty runs Claude Code and Codex inside real PTYs, then layers supervision and memory on top. A Rust daemon manages agent processes while a Next.js dashboard exposes terminals, memory controls, and a lightweight game-style control surface.

/Persistent PTY-backed agent sessions
/Shared project memory and per-agent notes
/Project-scoped MCP bridging between Claude and Codex
/Resumable sessions after daemon restarts
/Remote supervision through browser and Telegram

Houseparty

Why the architecture matters

The project is local-first on purpose. Repo access, CLI auth state, and execution stay close to the user machine rather than getting centralized too early. That makes the system more trustworthy today and creates a cleaner path toward a later control-plane plus runner architecture if it ever becomes a managed product.

The open-source readiness work also forced a useful distinction between a compelling internal tool and a broadly reusable product. Lifecycle tests, WebSocket behavior, MCP paths, and install ergonomics all matter once software leaves the author's machine.

Houseparty

Why it matters

The main lesson from Houseparty is that the missing infrastructure layer for coding agents is often statefulness and supervision rather than more model cleverness. Once the agent is treated as a durable session, questions about memory, permissions, worktree isolation, and operator intervention become much easier to reason about.