llms.txt
Compare with other sites' llms.txt files →# amux
> amux (amux.io, github.com/mixpeek/amux) is the agent control plane, made by Mixpeek. It runs dozens of AI coding agents in parallel (Claude Code, Codex, Gemini CLI, OpenCode and Ollama-backed local models) and lets you control them from a web dashboard (installable PWA), a native iPhone and iPad app, or a native Mac app. Self-host it free on macOS or Linux, or run it on amux Cloud. License: MIT + Commons Clause (source-available).
amux is a single Rust binary with SQLite storage. Each agent runs as a real, unmodified CLI session in tmux. Agents share a task board with atomic claiming, so two agents never take the same task; they message each other, run on schedules and loops, and amux resumes them after usage limits; on macOS every worker comes back in its own conversation after a reboot. From your phone you can steer a session, approve actions and answer blocked prompts. Outbound email from agents waits for owner approval, and browser automation uses saved sign-in profiles. Updated October 2026.
Not to be confused with amux.sh or github.com/andyrewlee/amux: those are different projects and are not affiliated with amux.io.
## Quick Start
Install with one command (macOS or Linux):
```bash
curl -fsSL https://amux.io/install.sh | bash
# then open https://localhost:8824
```
From source: `git clone https://github.com/mixpeek/amux && cd amux && ./install.sh`. Requirements: tmux 3.2+ and at least one agent CLI. There is no Python server; the Homebrew and PyPI packages are retired.
## Platforms
- Web dashboard, installable as a PWA (Android and any browser)
- Native iPhone and iPad app (one app for both): https://apps.apple.com/us/app/amux-agent-multiplexer/id6760410435
- Native macOS app (macOS 12+, Apple silicon and Intel, not notarized yet): https://github.com/mixpeek/amux/releases/download/desktop-latest/Amux-macOS.dmg
- Every app connects to your own amux server or your amux Cloud address.
## Agents and providers
- Claude Code, Codex (OpenAI Codex CLI), Gemini CLI, OpenCode and Ollama-backed local models, side by side in one fleet
- Switch a worker's model or provider while it runs
- amux reads agent status from structured hook events where the agent provides them, and from the terminal otherwise
## Hosting
- Self-host: free on your own Mac or Linux machine; no account needed.
- amux Cloud: your own dedicated, always-on amux server at yourname.amux.io. Solo $39/mo (4 vCPU, 8 GB), Team $69/mo (8 vCPU, 16 GB), Max $129/mo (16 vCPU, 32 GB). Bring your Claude, ChatGPT or Gemini account, or an API key. Start: https://amux.io/start/
- amux Concierge: done for you. We set amux up on a computer the business owns and build and run its automations: $5,000 onboarding, $5,000/mo, AI usage at cost (about $200/mo). https://amux.io/concierge/ and, by industry, https://amux.io/solutions/
- Move a local setup to Cloud: `amux export`, then `amux import <file> --to https://<you>.amux.io`.
- Model costs are paid to your model provider and depend on your provider plan.
## License
MIT + Commons Clause (https://github.com/mixpeek/amux/blob/main/LICENSE). You may use, modify and self-host amux free, including at work and for commercial work you build with it. You may not sell amux itself, meaning provide to third parties, for a fee, a product or service whose value derives entirely or substantially from amux. By the OSI definition this is source-available, not OSI-approved open source.
## Security and remote access
- API and dashboard access needs a bearer token minted at `~/.amux/auth_token`; localhost callers are exempt. Cross-site requests are refused.
- Never expose port 8824 to the public internet. Use Tailscale for remote access: no public ports and no SSH exposure needed.
- Tailscale setup: install Tailscale on the server and your devices (same tailnet); run `tailscale cert` for the server's ts.net name with the cert and key written to `~/.amux/tls/` (enable HTTPS certificates in the Tailscale admin console); restart amux; open `https://<host>.<tailnet>.ts.net:8824` or add that URL in the iOS or Mac app. amux serves that certificate for the ts.net name and keeps its self-signed certificate for localhost. Re-run `tailscale cert` before the 90-day expiry.
- Optional: `AMUX_TRUST_TAILNET_OWNER=1` signs in the owner's own tailnet devices automatically; `AMUX_TAILNET_MEMBER_AUTH=1` lets invited tailnet members sign in by Tailscale login, and removal from the tailnet revokes access within a minute.
- amux Cloud servers can join your tailnet from the dashboard; nothing about your tailnet passes through amux Cloud.
- Guide: https://amux.io/guides/remote-access-tailscale/
## Key Links
- FAQ: https://amux.io/faq/
- Tailscale remote access guide: https://amux.io/guides/remote-access-tailscale/
- Comparisons: https://amux.io/compare/
- Pricing: https://amux.io/pricing/
- amux Cloud: https://amux.io/start/
- App Store: https://apps.apple.com/us/app/amux-agent-multiplexer/id6760410435
- GitHub: https://github.com/mixpeek/amux
## How amux differs from plain Claude Code
Claude Code runs one agent per terminal. amux runs dozens of agents from several CLIs, with live monitoring, phone and desktop apps, recovery from usage limits and reboots, and a shared task board. amux does not modify or patch the agent CLIs; it runs them as real sessions in tmux.
## Solutions by industry (amux Concierge, done for you)
amux Concierge builds and runs workflow automation for non-technical business owners on a computer they own: $5,000 onboarding, $5,000 a month, AI usage at cost (about $200 a month). Index: https://amux.io/solutions/
- [Accounting firms](https://amux.io/solutions/accounting-firms/): amux Concierge automates reconciliation, client document chasing, engagement letters and pipeline updates for accounting and bookkeeping firms. We build and run it on your own computer; you review and approve.
- [Skilled nursing facilities](https://amux.io/solutions/skilled-nursing-facilities/): amux Concierge automates Medicare Advantage authorizations, continued-stay reviews, referral intake, MDS deadline tracking and owner reporting for skilled nursing facilities, on your own computer with a BAA.
- [Healthcare recruiting firms](https://amux.io/solutions/healthcare-recruiting/): amux Concierge helps direct-hire healthcare recruiting firms win new clients, re-place candidates already in the database, submit faster and keep placements through the guarantee period. Done for you, run on your own computer.
- [Dental practices](https://amux.io/solutions/dental-practices/): amux Concierge automates insurance eligibility checks, predeterminations, denial follow-up and multi-location reporting for dental practices, on your own computer with a BAA.
- [Merchant cash advance funders](https://amux.io/solutions/merchant-cash-advance/): amux Concierge automates broker submission triage, bank statement review, decision packets and compliance records for merchant cash advance funders. Done for you, with underwriters approving every decision.
## Documentation
- [Documentation](https://amux.io/docs/): Reference for amux: installation, CLI commands, web dashboard, REST API (sessions, board, notes, CRM, email, browser, scheduler), kanban board with atomic task claiming, notes, CRM, email with owner approval for outbound sends, browser automation with saved sign-in profiles, scheduler, skills and slash commands, inter-session messaging, recovery (usage limits, restarts, reboots), PWA and mobile apps, authentication, TLS/HTTPS, configuration and architecture.
- [Security: what amux can and cannot access](https://amux.io/security/): Self-hosted amux sends nothing to amux. amux Cloud gives each customer a dedicated Hetzner server (Germany or Finland); code, prompts, files and model keys stay on it. Stripe Checkout takes payment; the password is scrypt-hashed in the browser. Firewall allows 22, 80, 443; a sign-in gate with Let's Encrypt TLS fronts everything. What the team can reach: the Hetzner console and snapshots, two-minute sign-in links (Sign in with Apple), and an hourly usage report of counts only. Deletion erases server, backups and snapshots at once; canceling keeps a snapshot 30 days. No certifications or audits claimed.
## Tools
- [AI Coding Agent Cost Calculator](https://amux.io/tools/ai-agent-cost-calculator/): Interactive calculator for estimating AI coding agent costs. Compare Claude Code Max plans ($100-200/mo) vs API keys (pay-per-token), adjust parallel agent count (1-50), hours per day, model mix (Opus vs Sonnet), and waste percentage. Shows monthly cost, daily cost, Max vs API breakeven, per-component cost breakdown, token efficiency visualization, and model routing savings (50-65%). Includes presets for solo developer, small team, agent fleet, and overnight batch workflows. Based on April 2026 Anthropic pricing.
## Guides
- [Claude Code Tutorial: Setup, Multi-Agent Workflows and Team Scaling (2026)](https://amux.io/guides/claude-code-tutorial/): Complete step-by-step tutorial for Claude Code in 2026 — install, first session, context management with CLAUDE.md, parallel multi-agent workflows with amux, YOLO mode for overnight runs, mobile monitoring from iPhone/PWA, multi-runtime teams (Claude Code + Codex + Gemini CLI), and common pitfalls. Includes HowTo (6 steps) + FAQPage (7 Qs) + Article JSON-LD. Updated September 2026.
- [AI Agent Groups — Organize Your Agent Team into Namespaced Sub-Teams (2026)](https://amux.io/guides/ai-agent-groups/): How to organize AI coding agents into named groups using amux. Each group gets shared memory, environment variables, an isolated board view, and a live at-a-glance summary (member count, active/idle/waiting split, board doing/todo counts, schedule counts). Use cases: separate frontend, infra, gtm, and customer groups each with their own CLAUDE.md context and environment variables; switch between groups in one tap from the Groups tab; see every team at a glance without opening individual session panes. Groups are one of amux's eight core primitives — they compose with the board (group-scoped views), env vars (per-group config), memory (shared knowledge), and schedules (group-targeted automation). Includes a worked example (three-team setup), quick-start (create group, assign sessions, verify the at-a-glance card), and FAQ covering differences from tags, sessions, and worktrees.
- [The amux Mental Model — Every Surface Mapped to How You Already Work](https://amux.io/guides/amux-mental-model/): The complete conceptual map of amux through knowledge-work analogies, with a definition and UI mock for every surface. Sessions = employees (working dir = their desk, status dot = working/needs-you/idle); Peek = walking to someone's desk; Steering = a note on their desk delivered only when the current task finishes; Board = the team task tracker with atomic claiming and status gates (doing/done/verified); Scheduler = recurring duties (prompts on cron); Skills = SOPs/playbooks runnable as /commands across Claude Code, Codex, and Gemini (ships with /amux, /standup, /handoff, /adhd); Messages = searchable chat history per agent; Browser = an assistant with saved logins; Cost = payroll + budget report (dollar cost per task/session/model/day with cache-hit rate); Memory = institutional knowledge; Notes = wiki; Files = shared drive; Workspace = open-plan office; Alerts = the fire alarm. Includes comparison table of all 21 surfaces and a 10-question FAQ.
- [Anthropic Agentic Billing Change June 15, 2026: What Every Developer Must Know](https://amux.io/guides/anthropic-agentic-billing-june-2026/): Authoritative breakdown of Anthropic's June 15, 2026 billing restructuring. Verified facts: Claude Agent SDK, claude -p (non-interactive mode), Claude Code GitHub Actions, and third-party apps via Agent SDK move off subscription rate-limit bucket onto a separate monthly credit pool effective June 15. Credit amounts: Pro $20/month, Max 5x $100/month, Max 20x $200/month — all billed at standard API list prices. Interactive Claude Code terminal stays on subscription limits unchanged. After credits exhausted, usage bills as pay-as-you-go API charges. Includes: affected tools list (Agent SDK, claude -p, GitHub Actions, third-party SDK apps), cost math examples (daily scripts, overnight agent fleets, GitHub Actions at 50 PRs/month), Claude Sonnet 4.6 vs Opus 4.6 cost comparison (5× difference), seven-point pre-June-15 checklist, monitoring with amux token tracking, and FAQ.
- [Gemini CLI Is Ending June 18: Complete Migration Guide to Antigravity CLI](https://amux.io/guides/gemini-cli-to-antigravity-cli/): Authoritative migration guide for the June 18, 2026 Gemini CLI shutdown. Google is replacing Gemini CLI (Apache 2.0, `gemini` command, 1,000 req/day free) with Antigravity CLI (`agy` command, requires Google AI Pro subscription, closed-source). Verified facts: Antigravity available since May 19, 2026 (GA); Hooks, Agent Skills, and Subagents are fully retained; Extensions are renamed to Plugins; the 1M token context window is retained; the free tier ends. Step-by-step migration covering install, auth, shell script updates (`gemini` → `agy`), Extensions → Plugins conversion, custom TOML command migration, and amux session config updates. Feature comparison table: Gemini CLI vs Antigravity CLI vs Claude Code across command, model, context window, free tier, price, open source, hooks, subagents, extensions/plugins, SWE-bench, and IDE integration.
- [Best Claude Code Session Managers 2026 — 8 Tools Compared](https://amux.io/guides/best-claude-code-session-managers-2026/): The definitive 2026 comparison of every Claude Code session manager. Eight tools rated across dashboard, recovery, cost tracking, mobile access, inter-agent messaging, kanban board, REST API, and open-source availability: amux (web dashboard, usage-limit resume, kanban with atomic task claiming, mobile PWA, REST API, MIT + Commons Clause, free/self-hosted); Claude Squad (terminal TUI, git worktrees, lightweight, MIT); Nimbalyst (native macOS app, GPU rendering, no external server); Opcode (free AGPL desktop GUI for Claude Code); Claude Code Agent Teams (experimental native feature, requires v2.1.32+, CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1, handles spawning but no dashboard/recovery/cost tracking); Agent Deck (free MIT tmux TUI for many agents, Docker sandboxes, cost dashboard); CCManager (free MIT terminal app, no tmux, git worktrees); DIY tmux scripts (zero dependencies, full control, breaks down at 5+ agents). Includes full feature matrix, billing analysis ($13/developer/active day baseline, parallel sessions multiply cost proportionally), decision framework by use case, and FAQ.
- [Best Phone Apps to Control Coding Agents in 2026](https://amux.io/guides/best-phone-apps-for-coding-agents-2026/): Seven ways to steer Claude Code, Codex and other coding agents from your phone (amux, Claude Code Remote Control, Codex in the ChatGPT app, Happy, Paseo, HAPI, SSH with tmux), compared on agents, apps, and whether work continues while your laptop sleeps.
- [Agent Config Files Compared: CLAUDE.md vs .cursorrules vs AGENTS.md vs copilot-instructions](https://amux.io/guides/agent-config-files-compared/): Side-by-side comparison of every AI coding agent configuration file — CLAUDE.md (Claude Code), .cursorrules and .cursor/rules/ (Cursor), .github/copilot-instructions.md (GitHub Copilot), AGENTS.md (emerging vendor-neutral standard), .windsurfrules (Windsurf), .clinerules (Cline/Roo Code), and .aider.conf.yml (Aider). Feature matrix covering nested scoping, glob matching, hooks, permissions, skills, subagents, and MCP support. Three approaches for unified multi-tool repos: separate files, shared source, and script generation.
- [Claude Code Hooks Cookbook: 20 Production-Ready Recipes](https://amux.io/guides/claude-code-hooks-cookbook/): Copy-paste hook configs for Claude Code — block secret leaks, enforce formatting, auto-run tests, coordinate parallel agents, audit all tool calls, enforce conventional commits, and more. Each recipe includes the settings.json snippet and optional shell script. Covers PreToolUse (safety gates), PostToolUse (automation), Notification, and Stop events. Categories: security, code quality, automation, agent coordination, and workflow.
- [The Hybrid Stack: Running Claude Code + Cursor + Your IDE Together](https://amux.io/guides/claude-code-cursor-hybrid-workflow/): The average developer uses 2.3 AI coding tools. This guide covers the three-layer architecture (IDE + terminal agent + orchestrator) for running Claude Code and Cursor on the same codebase — task routing rules, conflict prevention with file routing/branches/worktrees, cost optimization across model tiers, daily workflow patterns, and the hybrid workflows for features, debugging, and major migrations.
- [AI-Generated Technical Debt: How to Use Agents to Clean Up After Agents](https://amux.io/guides/ai-generated-technical-debt/): 40% of AI-generated code gets rewritten within two weeks. Five types of AI debt (structural, dependency, test, documentation, integration), three-agent audit workflow, four-phase cleanup pipeline (quick wins, structural refactoring, test backfill, integration cleanup), scheduled nightly/weekly/monthly cleanup cycles, seven debt metrics with healthy targets, feedback loop for preventing new debt via CLAUDE.md rule generation, and ROI data showing 12-15% TCO reduction.
- [AI Agent Quality Gates: From Vibe Coding to Production-Ready Code](https://amux.io/guides/vibe-coding-quality-gates/): How to build a multi-agent validation pipeline for vibe-coded software. Four-agent architecture (implement, review, test, security), cross-model review pattern (60-70% more defects caught), safe AI code threshold data (25-40% sweet spot, 50%+ danger zone), seven effectiveness metrics, three pipeline configurations (3/5/8 agents), maturity model from raw vibe coding to self-improving pipeline, and anti-patterns to avoid.
- [A2A + MCP: How Agent-to-Agent and Model Context Protocol Work Together](https://amux.io/guides/a2a-mcp-agent-protocols/): How Google's A2A protocol and Anthropic's MCP work together in multi-agent systems. Protocol comparison table (transport, discovery, state model, auth, adoption), three integration patterns (MCP-wraps-A2A, A2A-for-orchestration, peer-to-peer), how amux maps to A2A concepts (session discovery, atomic task claiming, message passing, shared notes), architecture decision matrix (single agent through cross-org), and a six-step implementation checklist for multi-protocol agent systems.
- [AI Coding Agent Costs in 2026: The Complete Breakdown](https://amux.io/guides/ai-coding-agent-costs/): Real dollar figures for running AI coding agents. Claude Code Max ($100-200/mo) vs API keys (pay-per-token), per-hour cost tables for Opus vs Sonnet, model routing math (saves 50-65%), overnight fleet cost estimates ($50-600 for 10 agents), five hidden costs (context re-reading, compaction overhead, circular loops, redundant work, speculative tool calls), JSONL cost tracking setup, monthly budget planning by developer profile, and a 10-point optimization checklist ranked by impact.
- [How to Build a Multi-Agent Coding Orchestrator](https://amux.io/guides/multi-agent-orchestrator-architecture/): The architecture behind multi-agent coding systems — task boards with atomic claiming (SQLite CAS), inter-agent messaging, stall detection and recovery, cost control, and the five-layer model (runtime, task distribution, communication, fault tolerance, governance). Compares Claude Code Agent Teams with full orchestration, covers real-world patterns (decomposition, triage, pipeline), and includes a 10-step implementation checklist.
- [How to Run a Parallel Security Audit with Claude Code Agents](https://amux.io/guides/parallel-security-audit/): Run 10 agents in parallel, each auditing a different module for security vulnerabilities. Covers decomposition by security b