Choose your mode
Observe plans, approve proposed nudges, or explicitly enable bounded automatic nudging. You choose the authority.
OpenAGI✦ A personal autonomous agent
A self-improving,
proactive agent.
It learns. It reaches out. It earns its keep.
A personal agent that keeps context and keeps work moving. Chat across sessions, learn useful routines from activity you choose to share, and supervise your coding agents from one place. Connect your own models, tools and devices, with controls for what the agent can do.
curl -fsSL openagi.sh | shDownload the Mac appmenubar app · no terminal required
Explore Supervisor ↓ 5-minute setup·macOS, Linux, Docker, Raspberry Pi·Source-available · PolyForm NC·Your host. Your permissions.
openagi · illustrative examples
3 things to show you
[Mon · 7:42 am]
↗ I noticed a morning routine.
For the past 5 weekdays you:
· check #incidents in Slack
· pull yesterday’s deploys from GitHub
· draft a standup note
I can run this every morning at 7:30 am and have it waiting.
→ approved · saved as /morning-standup
[Mon · 10:14 am]
↗ Heads up — lunch with Sarah at 2 pm.
From your past conversations:
· her wedding was April 12 — ask how it went
· her dad’s surgery was scheduled for May 6
· she wanted feedback on her pitch deck
Want me to draft a few things to ask about?
[Mon · 2:31 pm]
↗ Prep for your 3 pm with Acme Corp.
Pulled via buildbetter-mcp:
· 4 tickets in last 30 days about CSV export limits
· mentioned competitor “Vellum” last call
· renewal in 6 weeks — expansion potential
Brief drafted. Want me to open it?
Supervisor
One view of your
coding agents.
See which Codex, Claude Code, Conductor and Antigravity sessions are working, waiting on checks, blocked, or waiting for you. Open the thread and PR behind each question, with verification evidence to help you decide what needs attention.
Set up Supervisor →Answer with context
Review the agent’s decision and its context. Answer through a supported route or open the agent’s app.
Talk through the queue
Optional voice on Mac, Android and Even G2 lets you ask about the fleet. Actions require a readback and spoken approval code.
Improve the system
Health metrics surface failures and repeated nudges. A configured repair worker can open and review a bounded PR for you to merge.
More ways to work
Your agent, where you need it.
Use the Mac app or a browser, and connect the tools and devices you choose. Availability varies by platform and configuration.
Mobile source / QA builds
Across your devices
Paired phones and compatible glasses bring tasks, chat, approvals and Supervisor along. iOS also has an offline workspace and a phone-agent mode on compatible devices. Pairing needs a reachable host; optional iOS alerts use encrypted payloads through a relay and Apple’s push service.
Mobile setup and availability →Connections in current source
Connect the work you choose
Connect Gmail, Google Calendar, approved messaging contacts, MCP tools and reviewed skills. Use your own accounts for optional messaging providers. Pebble Index transcripts can become memories and reminder tasks. Email and calendar writes require approval.
Explore connections →Latest safeguards in current source
Keep long runs bounded
Track model spend and set daily, per-turn and estimated input limits. When a limit stops a long task, saved progress helps you continue with a record of completed actions. Estimates are not billing guarantees.
Check the latest release →Mobile apps are source/QA builds, with no public store download promised. Connections and safeguards added after a release require a source update until included in a published build. Check the release notes and setup guides for your installation.
Privacy and control
Your host. Your models. Your permissions.
Your agent’s state lives on the host you choose. Cloud models receive selected prompts and tool context; connected services receive the data needed for their actions. Use a supported local model when you want inference on your device. Screen observation is optional, and live computer actions have separate permissions and approval. Native capture and computer control vary by platform.
vs. the rest
Local agents are good.
An agent with a decision layer is better.
Hermes Agent, OpenClaw, and SwarmClaw offer memory, MCP and self-hosted runtimes. OpenAGI combines optional observation, signal scoring, persistent specialists and saved corrections with Supervisor. Choose the model, permissions and connected tools that fit your work.
OpenAGI local | Hermes Agent local | OpenClaw local | SwarmClaw local | Operator cloud | Claude.ai cloud | |
|---|---|---|---|---|---|---|
| Runs on your machine | ✓ | ✓ | ✓ | ✓ | — | — |
| State on your own host | ✓ | ✓ | ✓ | ✓ | — | — |
| Bring your own LLM | ✓ any | ✓ 200+ | ✓ | ✓ 23+ | — | — |
| Watches you, learns patterns | ✓ | — | — | — | — | — |
| Adaptive Scrutiny decision layer | ✓ | — | — | — | — | — |
| Persistent specialists (propagation) | ✓ auto-spawn | parallel only | — | manual org chart | — | — |
| Saved corrections guide future work | ✓ | — | — | — | — | — |
| Persistent memory across sessions | ✓ tiered | ✓ FTS5 | ✓ markdown | ✓ durable | limited | limited |
| Multi-channel (SMS / Telegram / HTTP) | ✓ | ✓ | ✓ | dashboard only | — | — |
| Skills system | ✓ auto + manual | ✓ agentskills.io | ✓ | ✓ reviewed | — | — |
| Cron / scheduled tasks | ✓ | ✓ | ✓ | ✓ | — | — |
| Source-available | ✓ | ✓ | ✓ | ✓ | — | — |
| No telemetry, no accounts | ✓ | ✓ | ✓ | ✓ | — | — |
OpenAGI builds on the foundation that Hermes Agent, OpenClaw, and SwarmClaw shipped first — durable memory, MCP, daemon shape, BYO LLM. The four highlighted rows are the differentiators: watching you work, Scrutiny scoring, persistent specialization, and corrections that lock in. Hermes spawns parallel subagents per turn. SwarmClaw uses manually-configured org charts. OpenAGI auto-spawns persistent specialists when Scrutiny decides the task warrants one.
01 · Remembers
Talks like someone
who knows you.
Every conversation, every correction, every decision lives in OpenAGI’s tiered memory — short-term for working context, medium for repeated patterns, long-term “Lava” for durable truths. The agent doesn’t reset between sessions. It picks up where you left off, remembers what you ruled out, and stops asking the same questions twice.
The seven-axis Scrutiny layer scores every signal before acting. Corrections lock in. Mistakes don’t repeat. Risky or repeated tasks spawn bounded specialists with their own scope. Specialization without sprawl.
- → Tiered memory: short / medium / long-term, with intelligent decay
- → Directional Adaptive Scrutiny — act, ask, watch, ignore, propagate
- → Saved corrections help guide later decisions
Scrutiny · signal eval
urgency
0.82
impact
0.68
novelty
0.21
risk
0.14
confidence
0.91
specificity
0.76
conflict
0.05
Decision
act
Memory updated
→ lava (long-term)
act
ask
watch
ignore
propagate
02 · Watches
Killer featurePlus — it can
watch you work.
Turn on local screen capture and OpenAGI starts noticing the things you do over and over. The file you keep moving, the report you build every Friday, the script you run after every deploy. Once it sees the pattern enough times, it builds a skill for it and offers to do it for you next time.
You don’t teach it. You don’t configure it. You just work, and it’s paying attention. (And the screen capture stays on your machine like everything else.)
- → Local screen capture, processed on-device
- → Pattern detection across days, not just sessions
- → Auto-generated skills you can review, edit, or reject
- → Off by default — opt in per workspace
observed patterns · last 14 days
# patterns OpenAGI noticed without being asked
✓ /csv-to-s3
copy CSV from Downloads → rename → upload to S3 dev bucket
seen 6× · 14 days · confidence 0.84 · auto-skill ready
✓ /friday-report
pull metrics from PostHog → format → DM to #leads
seen 4× · 4 weeks · confidence 0.92 · auto-skill ready
○ unnamed pattern
screen recording → trim → drop in Slack thread
seen 3× · 7 days · confidence 0.61 · watching
03 · Yours
Personal. Private.
Yours.
Everything OpenAGI sees, learns, and remembers lives in ~/.openagi/ on your machine. No accounts. No telemetry. No remote dependencies beyond the model API key you choose to provide.
Use any LLM — OpenAI, Claude, Gemini, local Ollama, anything that speaks the OpenAI Responses API. Swap models any time. The agent you raise belongs to you, not to the lab that trained the weights.
- → Bring your own keys for any provider
- → Falls back to deterministic mode if no key is set
- → File-backed JSONL stores you can read, back up, or delete
- → MCP servers add tools without giving up control
~/.openagi
.openagi/
├── config.json # your model + keys
├── sessions/
├── memory/
│ ├── short.jsonl
│ ├── medium.jsonl
│ └── lava.jsonl # long-term
├── observations/ # patterns the agent saw
├── jobs.jsonl
├── skills/ # auto + handwritten
└── mcp/
# swap your model any time
$ openagi config set model=claude-opus-4
# or run fully local
$ openagi config set model=ollama:llama3
Architecture
Inputs in. Outputs out.
Everything in between is yours.
SMS
Telegram
HTTP
Cron tick
Screen capture
↓
Directional Adaptive Scrutiny
act · ask · watch · ignore · propagate
↓
Agent
Any LLM — OpenAI, Claude, Gemini, Ollama
Tools
Auto-learned skills, hand-written skills, MCP servers
Memory
Short, medium, long-term (Lava)
↓
Propagation
Repeated or risky tasks spawn bounded specialists with their own scope, memory, and tools. Specialization without sprawl.
↓
SMS reply
Tool call
Web UI
New skill saved
Install
One command. Or no command at all.
All three end with a daemon listening on 127.0.0.1:43210 and a setup wizard at /setup for your model keys.
curl -fsSL openagi.sh | shcurl
One-line shell installer. macOS & Linux. Drops the binary, sets up launchd or systemd, you’re done.
docker
Multi-arch (amd64 + arm64). Mount a data volume and the daemon survives restarts. Works on Raspberry Pi.
npm / source
Clone, run npm run serve, hit the setup wizard at localhost:43210/setup.
macOS app
A signed, notarized menubar app that bundles its own Node runtime — nothing to install first. Runs the same daemon, adds a Quick Ask panel on ⌥Space, screen capture, and auto-updates itself.
Raise your own agent.
Today, day one.
Start with one useful task, connect the context you choose, and expand what your agent can do as you gain confidence. Bring your own keys, your own machine and a model you trust.
$ curl -fsSL openagi.sh | sh