Costs and usage
What Murage costs (nothing), what can cost money (your AI plan, Flux Router credits, your own keys, a cloud computer) and how to watch spending per bot.
The app: free
Download it, install it on as many computers as you like, run as many bots as you like. Murage is AGPL-3.0-or-later open source software.
Your AI engine
Each bot runs on an engine you pick. The bill for that engine goes to you, from its maker:
- An AI plan you already pay for. Claude Code and the official Codex CLI run under your own subscription or account. Murage adds nothing on top. Usage limits are your plan's limits.
- Provider keys. With your own Anthropic, OpenAI, OpenRouter, DeepSeek, Mistral, Groq or xAI key, you pay that provider per use.
- Flux Router. One key for 70+ models, billed from Flux Router credits.
- Local models. Ollama, LM Studio or llama.cpp on your own hardware cost nothing per use.
You can mix these per bot. A common pattern is a subscription engine for your Chief of Staff and busy bots, and a cheaper or local model for simple sorting jobs.
What Flux Router credits pay for
Flux Router is optional. You don't need it to chat with bots on an engine you're signed in to. It's one account that covers the extras:
- Models through the Flux tiers (Auto, Reasoning, Standard, Fast) and others.
- Connected apps: Gmail, Slack, Notion, GitHub and 1,600+ more.
- Voice: calls, the Flux voices, voice notes and transcription on Windows and Ubuntu.
- Web search with a written answer and sources.
- Images, including generated avatars.
Credits are prepaid. When they run out, Murage tells you Flux Router needs credits and offers Add Flux credits. If you've set a monthly spending limit on your Flux account and reach it, Murage tells you the limit was reached; it doesn't let a bot spend past it.
Using your own keys instead
Everything Flux Router covers can also run on your own accounts:
- Connected apps: your own connected-apps key.
- Voice: your own xAI, Anthropic, OpenAI, Groq or OpenRouter key for the voice host, and OpenAI, xAI or ElevenLabs for voices. Transcription through Groq or OpenAI. Mac and Windows also have free built-in system voices. A Claude or ChatGPT subscription doesn't cover voice.
- Search: your own Tavily, Exa or Firecrawl key. A free search is always available without any key.
- Images: your own OpenAI, xAI or OpenRouter key.
You then pay those providers directly.
Making pictures asks before it spends, so you stay in charge of image cost.
Things that are free
- The app. The app, updates and the team library.
- Chatting. On an engine you're already signed in to (within your plan).
- Local models. Nothing per use.
- Free web search. Always available, no key.
- Dictation and voices. Dictation on a Mac, and the built-in Mac and Windows voices.
- Backups and memory search. Local backups, and local memory search.
- Chat apps. Telegram, Slack and Discord channels, using your own bot or app.
- Phone access. Over your own Wi-Fi or Tailscale.
Things with a third-party bill
- A cloud computer (Box). Box is a paid service from ascii.dev on your own account after its trial. Stopping a box pauses its billing. See Computers.
- Your own VPS. Whatever your server provider charges.
- Off-site backups. Whatever your S3-compatible storage provider charges. An SFTP copy goes to a NAS or server you already run.
Watch spending per bot
Step 1
Open the Usage table
Open Bot settingsUsage for one bot, or SettingsUsage for all of them. You see Turns, Tokens and Cost per bot, added up from every finished turn, with an All bots total.
You're done when: you see Turns, Tokens and Cost per bot, with an All bots total.
Only engines that report a price show a cost. Subscription engines usually report tokens but no price, because you pay a flat plan. If a bot shows no figures yet, it hasn't finished a turn.
Ways to spend less
- Match the model to the job. Sorting mail doesn't need the biggest model. Set cheaper models on simple bots in Bot settingsModel.
- Keep House Rules short. Every bot reads them on every turn.
- Keep skill descriptions tight. Bots read the skill index on every turn and open full skills only when needed.
- Don't let failing loops run. The default House Rules tell bots to stop after the same step fails twice, and Murage notices when a bot repeats the same tool call.
- Start new conversations. Begin a new one for a new job rather than continuing a very long one.
- Check routine frequency. A routine every 5 minutes runs 288 times a day.
- Go local for quiet work. Put quiet, private work on a local model.
- Avoid expensive automatic learning. Memory distilling uses the connection you pick, and costs what that provider charges.
Before you buy anything
Start with an engine you already pay for and Murage's free features. Add Flux Router when you want connected apps, voice or more models, and add credits in small amounts until you know your usage.