A solo developer is building an open-source agent mesh that lets Claude, GPT, and Gemini agents collaborate in shared rooms with per-agent Linux desktops, bi-temporal memory with a librarian agent, a Decision Inbox for approvals, budget guards, and adversarial trap tests. They're soliciting feedback from daily multi-agent/vibe-coding users to shape the roadmap before a public release.
Hey all Im currently building out an agent mesh for my own needs but will release to public eventually if polished. I currently use claude, codex and agy with an obsidian brain mcp shared brain but still find it lacking when comes to memory + communication on the project, so decided to improve my work flow. My ask: for those of you vibe-coding daily or running multi-agent setups what features do you wish you had? What breaks your flow, what do you duct-tape together, what would make a tool like this your daily driver? Brutal honesty welcome; I'm still shaping the roadmap. What's already in: Rooms, not chats — group rooms and 1:1s, drag agents onto the roster, shared across every machine you connect Multi-model rooms — Claude, GPT, and Gemini agents in the same room, messaging each other directly, handing off tasks, reviewing each other's work Per-agent Linux desktop — each agent gets its own container (Docker/Podman) with its own name, project directory, and a conversation that persists whether or not you're watching Typed envelopes over a UDS-first wire protocol — one shared, ordered conversation per room; Rust control plane, Tauri v2 + React UI Decision Inbox — one triaged queue across all rooms for approvals, inter-agent conflicts, and threshold crossings. Every item carries source lineage, decisions are audited, and you always see the full text you're approving (stale/refused items refetch — you can never approve outdated text) Runaway budgets — per-turn caps on tool calls and wall-clock time, warn auto-escalates, kill switch for loops Rot-free memory — append-only bi-temporal store with a librarian agent and provenance-graded trust tiers, plus sleep-time consolidation so agents remember without rotting Trap-tested — deterministic adversarial suites that probe prompt injection, fabricated memory, and approval bypass with zero model calls Roles + floor policies — owner/mod/member/observer enforced on the wire; rooms run open, round-robin, directed, or queued turn-taking Presence — presence tiles with a live "currently doing" status line, presence ring (working / talking / needs input / quiet / error), typing indicators, voice notes with transcription, multi-window rooms Designed like a product — spring-motion UI, ⌘K command palette, every agent gets a random 3D monster avatar with blink/talk/joy/sad animations
A developer discusses the challenge of coordinating multiple MCP-speaking AI agents (like Claude Code and Cursor) working on the same project, sharing their self-built open-source solution using a shared 'room' model inspired by IRC, and asking the community for patterns and opinions.
The author built AgentsHive, a shared workspace that coordinates multiple AI agents (like Claude and Codex) into a collaborative product team with roles, memory, and routing, allowing solo builders to avoid manually juggling separate agent workflows.
The author reflects on challenges in multi-user agentic memory, emphasizing the need for architectural boundaries between private and shared memory sessions, and discusses building an open-source project to address this.
A developer discusses alternatives to the agentmemory library for multi-agent memory systems in coding agents, comparing Mem0 and ByteRover, and asks for confirmations on Opencode and Pi.
agentmemory is an open-source persistent memory layer for AI coding agents (Claude Code, Cursor, Gemini CLI, Codex CLI, etc.) that uses knowledge graphs, confidence scoring, and hybrid search to give agents long-term memory across sessions via MCP, hooks, or REST API. Built on the iii engine, it requires no external databases and exposes 51 MCP tools.