Spawn→Work→Die→Remember

Your AI agents work alone. They don't have to.

Three open-source MCP servers that give AI agents persistent memory, voice I/O, and multi-terminal coordination. 45 tools. One ecosystem.

Voice→Claude→Memory

Claude→Orchestrate

BrainLayer

Persistent memory for AI agents

284K+ chunks indexed. 12 MCP tools. Semantic search, entity graph, knowledge digest.

Visit site →

VoiceLayer

Voice I/O for AI agents

540 tests. 11 MCP tools. Talk to your agents, hear them respond.

Visit site →

cmuxLayer

Multi-terminal orchestration

278 tests. 1,423x faster than shell. Spawn, coordinate, and monitor AI agents.

Visit site →

$ npx golems-cli install

Or install individually: pip install brainlayer · npm i voicelayer-mcp · npm i cmuxlayer

Get Started GitHub →

What Changes

Three problems. Three open-source solutions.

Your agent forgets everything between sessions

✓

Persistent memory across every conversation

BrainLayer

You type every command by hand

✓

Talk to your agents, hear them respond

VoiceLayer

One terminal, one task, waiting for each to finish

✓

Spawn parallel agents across coordinated panes

cmuxLayer

Get Started

One command for everything, or install what you need. Full guide →

Full ecosystem

$ npx golems-cli install

Installs all three MCP servers, wires Claude Code config, and verifies the setup.

or install individually

BrainLayerPython

pip install brainlayer

Persistent memory — semantic search, entity graph, knowledge digest

VoiceLayerNode.js

npm i voicelayer-mcp

Voice I/O — speech-to-text, text-to-speech, 11 MCP tools

cmuxLayerNode.js

npm i cmuxlayer

Multi-terminal — spawn agents, coordinate panes, monitor output

Full Setup Guide

45MCP Tools

2,557Tests Passing

295K+Knowledge Chunks

36Skills

3MCP Servers

Skills Library

36 reusable Claude Code skills. Install any skill with one paste.

Skills

With Evals

496

Assertions

94%

Eval Coverage

golems-cli skills install

$ golems-cli skills install cmux-agents

=== INSTALLING cmux-agents ===

Downloading: github.com/EtanHey/golems/skills/golem-powers/cmux-agents/

✓ SKILL.md (471 lines)

✓ evals/evals.json (3 evals, 18 assertions)

Symlinking: ~/.claude/commands/cmux-agents/ → installed

First-Time Setup:

Checking ~/.golems/config.yaml...

○ Config not found — running wizard

Detecting AI CLIs:

✓ claude /usr/local/bin/claude

✓ cursor /usr/local/bin/cursor

✓ gemini /usr/local/bin/gemini

○ codex not found

○ kiro not found

Workspace path: ~/Projects

✓ Config written to ~/.golems/config.yaml

✔ cmux-agents installed! Use /cmux-agents in any Claude session.

cmux-agents9 evals · 23 assertions · fixturesPortable

Install Prompt

Paste this into any Claude Code session to install cmux-agents:

Install and configure the cmux-agents skill for managing AI agents
in terminal panes. Download from github.com/EtanHey/golems and
symlink to ~/.claude/commands/cmux-agents/. If ~/.golems/config.yaml
doesn't exist, auto-detect installed AI CLIs (claude, cursor, gemini,
codex, kiro) and ask for workspace path. Then follow the First-Time
Setup section in the SKILL.md.

First-Time Setup

Auto-detects installed AI CLIs (Claude, Cursor, Gemini, Codex, Kiro)
Asks for your workspace path
Writes ~/.golems/config.yaml
Ready — use /cmux-agents in any session

/commit3 evals · 10 assertions

CodeRabbit review + conventional commit

/pr-loop7 evals · 25 assertions

Full PR lifecycle: branch → test → review → merge

/worktrees

Git worktree management (create, switch, cleanup)

/test-plan3 evals · 13 assertions

Generate test plan from requirements

/lsp

Code intelligence via Language Server Protocol

/coderabbit3 evals · 8 assertions

AI code review with multiple workflows

/critique-waves3 evals · 9 assertions

Parallel verification agents for consensus

/cmux-agents9 evals · 23 assertions

Spawn AI agents in terminal panes — Claude, Cursor, Gemini, Codex, Kiro

/cmux3 evals · 8 assertions

Terminal multiplexer management

/railway

Deploy, logs, restart, env vars for cloud worker

/1password3 evals · 10 assertions

Secret management, env migration, vault ops

/github3 evals · 9 assertions

Issues, PRs, checks, releases via gh CLI

/convex3 evals · 9 assertions

Schema, deploy, data import/export, troubleshooting

/golem-install3 evals · 10 assertions

Setup golems ecosystem with guided wizard

/brainlayer

Search, store, recall from 284K+ memory chunks

/research3 evals · 10 assertions

Deep web research with multi-source synthesis

/context73 evals · 9 assertions

Up-to-date library documentation lookup

/github-research3 evals · 9 assertions

Explore and document unfamiliar repositories

/catchup3 evals · 10 assertions

Full branch context recovery

/obsidian3 evals · 9 assertions

Search, read, write Obsidian vault notes

/content

Multi-platform content creation and publishing

/voice-sessions3 evals · 8 assertions

Voice-powered debrief, QA, practice sessions

/video-showcase3 evals · 12 assertions

Remotion-based product/project animations

/presentation-builder3 evals · 7 assertions

Workshop-method slide deck builder

/youtube-pipeline3 evals · 10 assertions

Extract knowledge from YouTube via Exa

/never-fabricate3 evals · 9 assertions

Mandatory verification before reporting results

/large-plan3 evals · 11 assertions

Multi-phase plan scaffolding and execution

/archive2 evals · 8 assertions

Archive completed PRD stories

/writing-skills3 evals · 10 assertions

Create and validate new skills

/skills3 evals · 12 assertions

Discover and list available skills

/coach14 evals · 63 assertions

Health, schedule, life planning with Whoop integration

/interview-practice3 evals · 13 assertions

7-mode practice with Elo tracking

/prd3 evals · 9 assertions

Generate Product Requirement Documents

/brave3 evals · 13 assertions

Browser automation via Brave

/figma-loop3 evals · 13 assertions

Iterative Figma-to-implementation workflow

/figma-swarm3 evals · 0 assertions

Multi-agent Figma screen decomposition and component build pipeline

/cli-agents3 evals · 13 assertions

Run Gemini, Cursor, Codex, Kiro for research

Browse All Skills on GitHub →

Skills Work on Any AI

3-layer adapter architecture · validated across 3 non-Claude CLIs

Layer 1

SKILL.md

Universal best practices with platform-agnostic fallbacks. Any AI can follow it.

Layer 2

capabilities.yaml

Machine-readable feature matrix. Routes tasks to the right CLI automatically.

Layer 3

adapters/

Exact flags, models, and syntax per CLI. One file per AI. 40-80 lines each.

Cross-AI Portability Eval — cmux-agents (Mar 2026)

CodexGPT-5.48/10Before universal fallbacks were added

Gemini2.5 Pro10/10Perfect — read fallback table correctly

KiroDefault9.5/10One buried detail missed, now explicit

Scores above are weighted rubric totals (partial credit for 4/5). The detail page → shows binary assertion pass rates (pass = 5/5 only).

Skills with adapter support

/cmux-agents /pr-loop /commit /coach

Built by Etan Heyman

AI engineer building open-source tools for AI agents.

Etan Heyman

AI Infrastructure & Developer Tools

Designed and built 3 MCP servers, 36 reusable skills, and an autonomous agent ecosystem from scratch. Every component is open-source, tested, and documented.

Portfolio GitHub LinkedIn