Personal agent

Goals & plans

Muse turns goals into plans, then advances them step by step. Sensitive steps wait for your approval.

Habits

Recurring things you want to keep doing. Check in when you do them - streaks count consecutive checked periods only, and a missed day honestly resets to zero. No freebies, no fake streaks.

Workforce

Agent teams and scheduled tasks, running in this tab. External actions need your approval.

Roster

Four built-in specialists. Add your own for work that needs a different specialty.

What agents can see and do

Each agent sees its subtask and teammates' outputs, not your chats, memory or keys. The Tools button shows whether web search, page fetch and calculator are enabled for each agent. Isolation is within this tab, not a separate machine.

A team file carries your custom agents and automations as JSON - share it and someone else's Muse can adopt the same crew.

Automations

Recurring prompts Muse runs on a schedule while this tab is open. Tab closed = orb asleep: missed runs surface as honest catch-ups on your return - they never run somewhere else.
Run a workforce from any goal card (Goals & plans → Workforce). The coordinator splits the goal into a dependency graph of subtasks, assigns each to a roster agent, runs independent subtasks in parallel waves, hands each agent its dependencies' output, then merges everything into one deliverable. Adapted from Eigent's open-source workforce, rebuilt for this browser.

Memory

Muse's context filesystem (vm://memory/...): durable personal facts in your encrypted VM - deduplicated, revised in place, retrieved by relevance (similarity x recency x usage), temporary facts expire on their own. Edit or forget anything, any time.

Tools & skills

What Muse can genuinely do from this browser - each tool labeled: real, sandboxed, key needed, or open-network.
Muse writes the code, gates it, sandbox-tests it - you approve before it joins the registry.

Prompt library

Prompts you reach for often - one tap drops one into the composer, ready to edit or send. Stored in your VM like everything else.

Web search

Live search through your own key. TinyFish's key is free with no card and also powers page fetches; Tavily has a free monthly tier; Brave has a free credit tier. The key stays inside your VM.

Monid

Monid is a catalog of hundreds of live data endpoints - web search, scraping, enrichment, monitoring. Paste a key from app.monid.ai/access/api-keys. Honest limit: Monid's API only accepts browser calls from its own domains, so from this static page the calls are CORS-blocked until Monid opens that up - the tools say so rather than failing silently. Outside the browser, the Monid CLI works fine.

Skills

Skills teach Muse domain workflows - a name plus the instructions it should follow. Built-ins can be toggled; yours can be deleted.

Open network mode

By default this VM only talks to your model and search providers. Open network mode lets Muse fetch arbitrary pages and reach MCP servers - it widens this tab's network policy until you turn it off. Off is the safer default.

MCP servers

Model Context Protocol servers over HTTP (streamable transport). Needs open network mode and a CORS-friendly remote server - stdio/local servers are unreachable from any web page, that is a hard browser limit, not a choice.

Studio

Mini-apps Muse builds for you, kept in your VM. Run them in an isolated sandbox frame, or download and open them anywhere.

Permissions

You decide what Muse may touch, and how much. Change or revoke any time. Scopes gate what Muse is allowed to plan and draft.

Audit trail

Everything Muse has done and plans to do. Complete, inspectable, yours.

Evolve

Recursive self-improvement - Muse drafts a proposal, hard gates evaluate it, you approve or reject. A static page cannot rewrite its own code, so an approved proposal exports as a patch bundle or hands off to your agent. It never self-applies.

Settings

Model, key storage and your Personal VM.
Every key - model, search, Monid, MCP server - is spent through one credential broker: it attaches the key itself, only to that service's own address, logs each use (and every refused destination) in the Audit trail, and keeps keys out of the rest of the code. It stops accidents and sloppy code paths; like any in-app guard, it is hygiene, not a hardware vault.
One JSON file with everything: chats, goals, habits, memory, workforce, automations, tools, studio apps, audit. Keys never leave the device - API keys, MCP keys and privacy-cloak originals are NOT in the file, so it is safe to store anywhere. Importing can merge into what is here (newer copy wins on conflicts) or replace everything - either way it shows exactly what the file holds first.
When on, every outbound model call - chat, goals, teams, automations, everything - is scrubbed first: emails, phone numbers and card/ID numbers are detected automatically, and anything you teach below is cloaked too. Each real value gets one consistent synthetic twin, so the model still reasons over coherent text. Replies are un-swapped before you see them, so you always read your real data. Twins live only in this browser's store - encrypted when a passphrase is set - and audit logs record counts, never values.

Open Muse is open source and runs entirely in your browser. With a passphrase set, the store is encrypted with AES-GCM using a key derived from your passphrase - forget it and the data is unrecoverable by design.