encrypted · local-first · LLM-independent
A fleet of AI agents, running on your machine.
Inner Droid is a desktop app for autonomous agents with strictly gated powers. Hand them email, files, a shell, connectors. Watch every move they make. Take any of it back.
Free for the whole beta. No card, no trial clock.
macOS, Windows, and Linux. Your machine, your keys.
Mission control for autonomous agents
You run a team of AI agents. Each one has a job, a set of tools it's allowed to touch, and a model it runs on. They work on a heartbeat while the app sits in the background. You drop in to chat, approve the risky stuff, and get on with your day.
A team, not a chat window
AI agents work in parallel, delegate to each other, and run scheduled duties without you babysitting them.
Always-on, in the background
Awake agents act on a heartbeat: polling mail, running routines, picking up delegated work while you're elsewhere.
You stay in the loop
Anything sensitive stops and waits for you. Approve it once, or grant a standing rule when you trust it.
Put the fleet to work
Eight examples of what you could hand off. None of these come preinstalled: they're flows you build by describing them to Droid. Each one runs on a heartbeat in the background, and anything sensitive still stops at the gate.
usecases.idx
Inbox triage on autopilot
An agent polls your mail on a heartbeat, sorts what matters, and drafts the replies. Nothing leaves without your approval — you clear a queue, not an inbox.
usecase.01
Inbox triage on autopilot
An agent polls your mail on a heartbeat, sorts what matters, and drafts the replies. Nothing leaves without your approval — you clear a queue, not an inbox.
And these eight are just a sample. Describe the flow you need to Droid in plain conversation, and it wires up the agents, tools, and actions to run it.
Every AI agent is its own operator
An AI agent is an identity (its SOUL), a set of procedures, recurring routines, an allowlist of tools, a model, and a memory that sticks around. Some stay awake and act on every heartbeat. Others sleep until a teammate wakes them.
agent.def
- SOUL
- who the agent is
- INSTRUCTIONS
- how it works
- ROUTINES
- what it does on a schedule
- TOOLS
- the exact powers it's allowed
- MODEL
- the brain it runs on
- MEMORY
- what it remembers
Awake
Acts on every heartbeat. Polls, runs routines, delegates.
Asleep
Dormant until a peer hands it work.
Droid runs the fleet
A built-in AI agent named Droid is the one you talk to first. It walks you through creating new agents in plain conversation, routes work across the team, supervises everyone on a heartbeat, and picks up the slack as a fallback worker.
The tools come built in
Your AI agents don't start from zero. Memory, mail, calendar, and messaging are wired in from the first launch, each one behind the same gate.
Persistent dynamic memory
A knowledge base your AI agents build and keep current, not you. Point them at sources, ask questions, and they file what matters into plain, linked notes you can open and edit. Ask to recall something later and it comes back, instead of getting worked out from scratch every time.
ingest / recall / cross-link / index / log / lint
Smart email client
A full inbox your AI agents read, triage, draft, and send from. An agent polls it on a heartbeat; you approve anything that leaves.
Calendar and reminders
AI agents schedule, watch, and act on your calendar. Reminders fire on time and turn into work, not another notification you ignore.
Messaging connectors
Enable a channel and your agents read what comes in and reply through it. A customer messages you on WhatsApp; an agent reads it, pulls what it needs from memory, and answers. Your devs post updates in Slack; an agent follows the thread and chimes in.
WhatsApp · Discord · Telegram · Slack
Meeting recorder
Record a call from your microphone and the system audio, so the other people on it land in the transcript too. Whisper writes the markdown on your machine and the audio never leaves. No agent reads those transcripts until you hand it the folder.
record / transcribe / markdown / local
Need a tool that isn't here?
Chat with Droid and answer a few questions. It builds the custom tool for you and hands it back ready to use, behind the same gate as everything else.
describe / answer / generate / ready
The gate is the point
Every sensitive action runs through a policy engine before it happens. A shell command, a file it wants to delete, an email about to go out. It stops at the gate, lands in the approval ledger with a #ref, and waits for you.
Two-layer policy
Embedded rules you can't override sit above the ones you write. Safe defaults come first.
Attributable and reversible
Every action carries a #ref you can trace and undo from the interface. Nothing important happens silently.
Autonomy you grant
It starts locked down. Hand out an 'always allow' when an AI agent earns it, and take it back whenever you want.
approval ledger
- #482shell: git push origin mainscoutpending
- #481fs: delete ~/exports/oldarchivistapproved
- #480mail: send to 3 recipientsdroidheld
Sovereign by construction
The trust model isn't a setting you switch on. It's how the app is built.
Encrypted at rest
The whole database is encrypted on disk. An optional PIN gates the app itself.
LLM-independent
Run on your Claude subscription or your own OpenRouter key. Switch backends in settings. Each AI agent picks its own model.
No terminal exposed
Tools, policies, credentials, prompts: all of it lives in the app. Nothing leaks to a shell you forgot to close.
Local-first
It runs on your machine. Your data doesn't take a trip to someone's cloud to get the work done.
How it compares
OpenClaw and Hermes Agent are strong open-source harnesses for people who enjoy running their own infrastructure. Inner Droid puts the same kind of agent fleet inside a sealed desktop app: encrypted at rest, every sensitive action behind an approval gate, and nothing to manage from a terminal.
| Aspect | Inner Droid | OpenClaw | Hermes Agent |
|---|---|---|---|
| Getting started | One desktop app. Agents, policies, credentials, and approvals all live in the app; no terminal exposed. | Shell or npm install, plus a gateway daemon you keep running; the companion apps sit on top of a CLI core. | Desktop installer or CLI; the installer pulls in Python and Node tooling, and messaging needs a separate gateway daemon. |
| Action control | Every sensitive action stops at the gate with a #ref in an append-only ledger. Grants are per agent, revocable any time. | Exec approvals with allowlists and DM pairing; a per-action approval ledger is not documented. | Smart or manual command approvals and DM pairing; a unified action ledger is not documented. |
| Data at rest | Whole database encrypted on disk (SQLCipher); optional PIN backed by the OS keychain. | Plain local files; the docs point to file permissions and whole-disk encryption you set up yourself. | Plain local files; the docs rely on file permissions, with no at-rest encryption documented. |
| Privacy | Local-first: only what a tool explicitly sends ever leaves your machine. | Local-first with no vendor cloud; what a community skill sends is up to each skill. | Local-first with no telemetry; the optional Portal's data-retention policy is not documented. |
| Multi-agent | A native team: awake and asleep agents on a heartbeat, a built-in orchestrator, delegation, and a shared board. | Channel routing to isolated agents, plus subagents; each one is a workspace you configure. | One central agent that spawns isolated subagents for parallel work. |
| Built-in tools | Mail client, calendar, semantic memory, and WhatsApp / Slack / Discord / Telegram, preinstalled and all behind the gate. | A vast catalog — dozens of channels, browser, voice — assembled from plugins and community skills, including skills the agent writes for itself. | Sixty-plus tools: browsing, voice, cron, MCP, several sandbox backends; powerful, and yours to configure and secure. |
| LLM backends | Claude subscription or your OpenRouter key, switched in settings; each agent picks its own model. | Bring your own key, subscription, or local model, wired up in config. | Bring your own key or local models; the bundled route runs through the paid Nous Portal. |
| Cost & upkeep | Proprietary app in private beta; one vendor builds and maintains it. | Free (MIT), stewarded by a foundation; you are the operator, with updates and security as part of the job. | Free (MIT) to self-host; the optional Nous Portal is a paid subscription. |
Comparison drawn from each project's public documentation, July 2026. Details may change as the projects evolve.
Get in early
Inner Droid is in private beta. Leave your email and I'll send one message when it opens. That's the whole deal.
Free for the whole beta
Every build and every agent, from the day you get in until the beta closes. No card, no trial clock.