A command deck for your own agent. Replaces Martin and the rest of the $20–40/month "personal AI OS" tier. Part of Openware — software you own outright.
Talk to it. It answers out loud, one sentence at a time, across a caption line you can read from the other side of the room. It runs tools, remembers what you tell it, and stops to ask before it touches anything real. No account, no credits, no per-minute transcription bill.
The brain is yours: attach a self-hosted Hermes Agent and freejarvis becomes its face — its sessions, cron jobs, skills, toolsets and approvals, all live. Attach nothing and it runs its own loop against any OpenAI-compatible model.
git clone https://github.com/openwarehq/freejarvis
cd freejarvis
cp .env.example .env # add a model — free key, no card, about a minute
docker compose up
# http://localhost:4333
Read this part first.
- Speech recognition is the browser's, and in Chrome and Edge that means the audio goes to
Google. This is the one place in the drop where anything leaves your machine, it only
happens while you are holding the talk key, and it is worth knowing before you hold it.
Safari recognises on-device. Firefox has no recognition at all and the deck falls back to
typing. Text-to-speech is local in every browser. If you want fully offline voice, this is
the piece to replace, and it is one hook —
src/hooks/useVoice.ts. - There is no wake word. Hold space, or click the mic. Always-on listening means an open microphone and a recognition stream running all day, which is a different product and a different privacy conversation.
- ElevenLabs is optional, and it is a paid service. freejarvis speaks through your browser
by default — free, offline, no key, and it sounds like it. Give it an
ELEVENLABS_API_KEYand it sounds like a product. That key is a bill you pay ElevenLabs per character, and it is the one place a subscription can get back into this drop. It is opt-in, it is never required, and nothing degrades without it except the timbre. There is still no freejarvis account. - On the browser voice, the orb's motion while speaking is synthesised, not measured.
speechSynthesisexposes no waveform, so the amplitude comes off word-boundary events with a decay between them — close, but not the speech. On ElevenLabs the audio is a real stream through an AnalyserNode, so the orb is the actual voice. - The site viewer displays sites; it does not build them.
open_siteframes an.htmlfile that already exists in the folder you pointedSITES_DIRat. If you want a model to write the site, that is unlovable. - It does not read your inbox, run your calendar, or make phone calls. Martin does, and that is a real difference. freejarvis drives an agent — give the agent those tools, which is what Hermes skills are for, and it can. Out of the box it has eight tools, and none of them is your email.
- The scheduler lives in the web process. Nothing fires while the container is down, and a job interrupted mid-run does not resume — it runs again at its next scheduled time. A queue that needs its own service would be a second command, and the bar here is one.
- One brain at a time. Setting
HERMES_URLtakes precedence; the direct model is what you get when no agent is attached. There is no routing between them. - The deck has no login. It binds to localhost and assumes the only person who can reach port 4333 is you. Do not put it on the open internet without something in front of it.
- Hermes panels only show what your Hermes build actually serves. Toolsets, skills and jobs
are read from
/v1/toolsets,/v1/skillsand/api/jobs; an older build without one of those gets a panel that says so, rather than an empty table that reads as an agent with nothing in it. - Memory is a list with substring search, not embeddings. For a few hundred facts about one person that is the right amount of machinery. It is not a RAG system.
- Tool calling is only as good as your model. See Tested models below — the free tier default works, and it is not equally reliable at everything.
Two brains, one face. The deck does not know which one it is talking to. Both emit the same
event vocabulary — assistant.delta, tool.started, tool.completed, approval.required —
so every panel, the caption line and the orb work identically either way.
| Attached to Hermes | On its own | |
|---|---|---|
| Chat | Hermes' agent loop, its memory and skills | freejarvis' loop |
| Sessions | Hermes sessions + a local mirror | local SQLite |
| Cron | Hermes jobs, which can deliver to Telegram or Discord | in-process scheduler |
| Tools | whatever your agent has | eight built-ins, plus open_site and shell when enabled |
| Skills | yours | — |
| Approvals | Hermes run_approval |
local gating |
Voice both ways, and hearing costs nothing ever. Hold space, talk, let go. Recognition is the browser's — no Whisper key, no minutes, no transcription bill, and there is no paid tier of this that listens better. Speaking is the browser's too until you hand it an ElevenLabs key, at which point it starts on the first finished sentence rather than the last one, and fetches the next clip while the current one plays so the queue never gasps.
Ask for a site and it appears. Point SITES_DIR at a folder of .html files and every one
of them becomes something you can ask for out loud. The available names are compiled into the
tool's own description, which is the part that makes "pull up the Azur site" resolve instead of
guess. It opens windowed in the top-right corner so the orb stays on screen next to it —
expanding to the full frame is a deliberate click — and the chat dock stays live either way, so
you can ask for the next thing without closing this one.
The orb is a readout, not decoration. Five thousand points on a sphere, and every property is bound to something real — spin rate and displacement to what the agent is doing, the radial wave to microphone amplitude, and the hue of the entire interface to the current state. Standby is teal. Thinking is violet. Working is green. Waiting on you is red.
It stops before it does anything real. write_file and schedule pause the run — genuinely
pause it, the stream stays open — and show you the exact arguments. Approving write_file
tells you nothing; approving write_file {"path":"notes/today.md"} tells you everything.
Cron that reads like English. 0 8 * * 1-5 renders as weekdays at 8:00 am, with the next
firing time. Scheduled runs have nobody in the room, so gated tools are refused rather than
queued for an approval that will never come, and the answer lands in the activity feed.
It boots like something. Every line of the start-up sequence is a real check against a real endpoint — brain reachable, tools loaded, jobs armed, memories stored, voice engine. It looks like a title sequence because the numbers happen to be interesting. If the agent is unreachable, this is where you find out, and the line goes red and stays.
The readout is measured. Latency is wall-clock from send to first token — the number you actually feel — not a token count dressed up as speed.
SOUL.md, borrowed on purpose. Hermes keeps an identity file at ~/.hermes/SOUL.md, so
anyone arriving from there already knows what this file is. Edit it in the app; it is the whole
personality on a direct model and layers on top of your agent's own when one is attached.
Hermes ships an OpenAI-compatible API server. It is off by default. In ~/.hermes/.env:
API_SERVER_ENABLED=true
API_SERVER_KEY=$(openssl rand -hex 32)Then run the gateway, which prints the port it is on:
hermes gateway
# [API Server] API server listening on http://127.0.0.1:8642Point freejarvis at it — in the Settings panel, or in the repo-root .env:
HERMES_URL=http://localhost:8642
HERMES_KEY=<the same API_SERVER_KEY>You do not need to configure CORS. Hermes does not enable browser CORS by default, and
every call from this deck goes through its own server rather than from the page — which also
means API_SERVER_KEY never reaches your browser. localhost is rewritten to
host.docker.internal inside the container, so the same line works in both places.
Already running Hermes in Docker? Add API_SERVER_ENABLED=true, API_SERVER_HOST=0.0.0.0 and
a key to its environment, and point HERMES_URL at the container.
| Hold space | talk; release to send |
| Enter | send what you typed |
| Esc | stop the voice, then stop the run, then close the panel |
| ⌘⏎ | approve the pending tool |
| Esc (on the card) | deny it |
The left rail is Sessions, Memory, Cron, Tools, Skills, Approvals, Soul and Settings. The spheres in the distance are your other open sessions — one hue each, stable across reloads, so a conversation keeps its colour.
Eight built-ins on a direct model, and two more you switch on. Two of them ask first.
| Tool | ||
|---|---|---|
now |
the clock | it has none otherwise |
remember |
store a durable fact | |
recall |
search what it stored | |
read_web |
fetch a URL as readable text | no key |
list_files · read_file |
inside data/workspace |
|
write_file |
inside data/workspace |
asks first |
schedule |
create a cron job | asks first |
open_site |
frame a site on the deck | only when SITES_DIR is set |
post_reel |
carry the next clip through Instagram's uploader | stops before Share |
post_reel is the one tool that leaves the machine, and it is deliberately not gated. The
approval card is for acts that reach the outside world, and this one stops a button short of
one: it takes the oldest clip in the reel folder that has not gone yet, measures it, writes a
caption, opens a browser and fills the whole post in — then leaves Share alone. The gate is the
button, and the button is yours. See Sending a reel below.
Every path resolves and is compared against the real workspace root after symlinks — a link
planted in the workspace pointing at /etc does not work, and neither does ../.
There is one more, shell, and it is off:
FREEJARVIS_SHELL=1On, it runs commands in the workspace as the container user, and it still asks before every one. A web page that can run commands on your host is a different threat model than a web page that can read a file, and the difference should be a line you wrote on purpose.
http://localhost:4333/?demo=1&script=yacht # greet, status, open the listing
http://localhost:4333/?demo=1&script=portfolio # three sites, swapped in place
http://localhost:4333/?demo=1&script=reel # or just press ⌘⇧E anywhere on the deck
A scripted take, for filming. The deck runs its real components — the orb is the orb, the voice really is ElevenLabs, the site really is the file on your disk in a real iframe. The only thing that comes off the script is what the agent decides to do, and every step in it is something the agent genuinely does when you just ask it.
It exists because a free-tier model returns an empty response about one turn in ten and takes a variable two to six seconds to think. That is survivable when you are working and fatal when you are on take nine.
It records clean. No controls, no chat dock, no ticker — nothing to crop out. A hint card names the chords and removes itself after five seconds.
| ⌘⌥⇧A | fire a take (resets first) |
| ⌘⌥⇧R | reset |
| ⌘⌥⇧D | show/hide the controls |
Chords rather than single keys, because during a take the deck is listening to the room and a stray keypress is a ruined shot.
The client owns the timing, not the server. An earlier version paced the stream to each line's estimated spoken duration so that visual events landed on the right word. Estimating from character count is off by 15–27% on a short sentence, and every one of those errors is either a hole in the audio or an event firing while the voice is three lines back. So the whole script streams as fast as it renders, every clip is fetched the moment its sentence completes, and the speech queue plays them back to back — measured at 9–89 ms between clips. The site opens as a queued action between two real clips, so it lands on the line that announces it rather than on an estimate of it.
- Scripts live in
src/lib/demo.ts— plain data, edit freely - A
sayline can use{tools}{memories}{jobs}{sessions}{voice}, filled in from the live system at stream time — so the status report it reads out is the actual status report. Delete a job before filming and the line changes - Length is a character budget, not a step count: the voice reads about fourteen a second, so the shipped take is ~170 characters and runs ~12 seconds
- Prefer an em dash to a full stop where you want a beat — a second sentence is a second clip, and clips are where seams live
- Put a
toolstep before the line that describes what it did, not after. The action is queued into the speech timeline, so the site is already on screen while it is being named
Nothing reaches it by accident. And if someone wants to know whether this actually works, close it and just use it — that is the better demo.
⌘⇧E. That is the entire interface.
⌘⇧E
├─ oldest clip in ~/Movies/Freejarvis that has not gone yet
├─ 0:09 · 1080×1920 portrait · 41 cuts · 60fps
├─ a caption
├─ the browser opens — visible, you watch it
├─ Create → file attached → Next → Next → caption typed
└─ stops. Share is yours to press.
Everything else on this deck is something you talk to. This is the one thing you want done while your hands are somewhere else, so it is one key, off the ⌘⌥⇧ chord the filming keys use, and it works whether or not a take is loaded.
It runs as a script, and the tool step in that script is live — the rest of a take is a tape because a free-tier model is slow and occasionally empty, but a reel take that narrates an upload without uploading is not a demo of anything. That one step really opens the browser, really attaches the file, and really stops at Share.
brew install ffmpeg # the measurements and the stills come from it
npm run dev # the deck, on 4333
npm run reels:login # log into Instagram, once
mkdir -p ~/Movies/Freejarvis # drop clips hereThen ⌘⇧E.
reels:login matters more than it looks. Chrome allows one instance per
profile, so logging in with a plain --user-data-dir leaves a window holding
the profile without the debugging port — and the next ⌘⇧E cannot start its own,
so it waits twenty seconds and fails with nothing useful to say. Opening it with
the port from the start means the window you logged in with is the window the
deck drives. Nothing to quit, nothing to remember. If a portless window is
holding the profile anyway, it now says so immediately instead of timing out.
Its own Chrome profile, at data/reels/chrome.
Not a workaround. Since Chrome 136 the remote-debugging port is refused outright on the default profile, because "any local process can drive the browser you are logged into everywhere" is a real hole. A separate profile holds one session, for one site, that you put there on purpose — nothing here can reach your mail or the tabs in your other window.
Buttons are found by what they say. Instagram's class names change without notice;
._acan._acap is a selector with a shelf life measured in weeks, while "Create", "Next" and
"Share" have not moved in years and are how a person finds them too. The file never goes near
the OS file picker — CDP sets it on the <input type=file> directly. The caption goes in
through the input pipeline, because setting textContent on their composer looks right on
screen and posts empty.
With ANTHROPIC_API_KEY set, four stills are pulled evenly across the clip and a model writes
the caption from what is actually in frame, under instructions never to invent a place, a
person, a brand or an event.
Without a key it is built from the measurements and says only those:
hangar demo
0:09 · 41 cuts · 1080×1920 · 60fps
Shot on a Saturday, some time before dawn.
That is a style, not an apology. A caption that invents what is in a video nothing has looked at
is worse than a caption that counts. Screen Recording 2026-04-24 at 5.04.01 PM is not a title
and is thrown away rather than dressed up; my_edit_v2 is your own words and is kept. The cut
count comes from running the clip through ffmpeg's scene filter — nothing else in the metadata
tells you whether nine seconds holds one cut or forty.
~/Movies/Freejarvis, or wherever REELS_DIR points. Oldest first, so it is a queue and not a
stack. A file still being written is not ready — an AirDrop or an export lands as a growing
file, and picking one up halfway gives a corrupt upload with no error anywhere. What has flown
is remembered by content rather than path, so filing a clip into an archive afterwards does not
make it look new. And cancelling leaves the clip at the front of the queue: only a real post
is recorded, which is what you want when you cancelled on purpose.
Needs ffmpeg on the host (brew install ffmpeg).
In .env — copy .env.example and fill in one of the two brains:
LLM_BASE_URL=https://openrouter.ai/api/v1
LLM_MODEL=google/gemma-4-26b-a4b-it:free
LLM_API_KEY=Point LLM_BASE_URL at https://api.anthropic.com and it switches wire format itself —
x-api-key, the version header, the system prompt hoisted out of the message list, tool
results as content blocks. You configure none of that.
Everything else is optional:
HERMES_URL · HERMES_KEY |
attach an agent; takes precedence over LLM_* |
ELEVENLABS_API_KEY |
a voice worth listening to. Paid, opt-in, never required |
ELEVENLABS_VOICE_ID |
default onwK4e9ZLuTAKqWW03F9 — Daniel, British, steady |
ELEVENLABS_MODEL |
default eleven_turbo_v2_5; eleven_flash_v2_5 is faster and thinner |
SITES_DIR |
a folder of .html files the agent can pull up by name |
FREEJARVIS_SHELL=1 |
enable the shell tool |
PORT |
default 4333 |
DATA_DIR |
default ./data — database, SOUL.md, workspace |
Anything set in the Settings panel is stored in the database and beats the environment, so you
can hand this to someone who has never opened a .env.
Measured, not assumed. All over OpenRouter unless noted.
| Model | Chat | Tools | Notes |
|---|---|---|---|
google/gemma-4-26b-a4b-it:free |
✅ | ✅ | The repo default, and what every screenshot here was taken against. Calls tools correctly, including chained calls and the approval pause. Intermittently returns a completely empty response — perhaps one turn in ten, more often under a long SOUL.md. freejarvis names that instead of going quiet; send it again. |
Anthropic claude-* |
— | — | The wire adapter is written and unit-tested. No request has ever been sent to api.anthropic.com from this repo, so it is untested against the live API. |
| Ollama, LM Studio, vLLM | — | — | Should work — it is the same OpenAI shape — but tool calling depends on the model you pull, and small local models are the least reliable at it. Not measured. |
The deck degrades honestly: a model that will not call tools still chats, and the tools panel still lists what it declined to use.
freejarvis/
├── src/lib/
│ ├── events.ts the one event vocabulary both brains speak
│ ├── runner.ts picks a brain; nothing above this line knows which
│ ├── agent.ts the standalone loop — tools, approvals, memory
│ ├── hermes.ts the Hermes client, feature-detected
│ ├── llm.ts OpenAI + Anthropic streaming, with tool calls
│ ├── tools.ts eight built-ins and the workspace jail
│ ├── cron.ts a five-field cron parser
│ ├── sentences.ts stream → whole sentences, for the caption and the voice
│ ├── voice.ts ElevenLabs, server-side so the key stays here
│ ├── sites.ts the sites folder, and the jail around it
│ └── scheduler.ts the in-process tick
├── src/components/
│ ├── Orb.tsx the sphere
│ ├── SiteView.tsx a site, framed on the deck
│ ├── BootSequence.tsx real checks that happen to look cinematic
│ └── panels/ one per rail item
└── data/ your database, SOUL.md and workspace — gitignored
npm run dev # http://localhost:4333
npm test # 53 tests
npm run typecheckMIT — take it, sell it, host it, fork it.
Not affiliated with Nous Research, Martin, or anyone else named here. Hermes Agent is MIT and belongs to Nous Research; this is a separate program that talks to its public API.