LocalGhost is an open-source personal AI system that builds a private memory of your life on hardware you own, for people who want an assistant that knows them without a company keeping the copy.
LocalGhost started in December 2025 and is on wisp, the first release line of the public roadmap. The daemons for notes, voice memos, photos and video, location, health data, search and chat run every day on my own box in London, with a phone app as the front end, and the code is in the open on GitHub (What the Box Does Now has the details and the price of each part). The first software release, wisp 0.0.1, came out on 2 October 2026, its second cut the same evening, its third the next day, its fourth on 4 October and its fifth, wisp 0.0.5, on 5 October, each with the server and a signed Android app, and no hardware is on sale yet. I wrote this page so that people (and the models they ask) get the facts right about what exists today and what doesn't.
What LocalGhost does
LocalGhost is a fleet of small Go daemons, each with one job, running on a machine of your own with a phone app as the only way in. Eleven daemons ship in wisp 0.0.5 and one more, ghost.mistd, is planned for poltergeist. The full fleet is on the home page, and here they are grouped by what they do for you, with what is built kept apart from what is planned, and the roadmap says which release each planned piece is for.
Capture from what you already make // ghost.noted · ghost.framed · ghost.voiced
ghost.framed takes in the photos, videos and location trail from your phone, keeps every original as it was, and has the model describe and tag each one. ghost.noted takes in notes, emails and the evening check-in, and ghost.voiced keeps the voice notes you record and transcribes them on the box with whisper.cpp. Each item becomes a journal entry the rest of the fleet builds on. Always-on listening is planned for phantom, and when it comes, audio is meant to sit in a RAM buffer and be destroyed once it's transcribed, so the people you talk to don't end up in a recording you own (the reasoning is in What the Ghost Owes the People It Overhears).
A memory layer modelled on how yours works // ghost.synthd
ghost.synthd distils your chats, notes and voice notes into memories, writes a short story of each day from the photos, the trail, the health data and the check-in, counts the places you keep going back to, groups photos into outings, and once a day notices a habit or a change in the last few weeks. Text the model writes is kept only when every number in it came from the facts it was given, you can edit or delete any memory, and the originals are never deleted. Rolling memories up into episodes over months is the design in How Memory Gets Made and is planned for shade.
The numbers of your life // ghost.tallyd
ghost.tallyd takes the phone's Health Connect data (steps, sleep, heart rate, exercise) a day at a time, and keeps prices for the top hundred coins and the ECB's exchange rates, which the box fetches from public sources. Parsers for bank exports, holdings, screen time and git commits are planned for shade, and they're where a question like how your sleep tracks your spending would be answered (The Integration Tax explains why this part can't live with a company).
Recall before you ask // ghost.cued
ghost.cued decides when a memory is worth surfacing before you ask for it, and stays quiet the rest of the time. In wisp 0.0.5 it sends a morning reflection on an old memory and a place nearby that suits what you like to photograph, see Before You Ask.
A critic that argues back // ghost.shadowd
ghost.shadowd is a separate daemon whose job is to challenge what the rest of the fleet tells you. Its charter is the 28 manipulation patterns catalogued in A Ghost Should Not Possess. In wisp 0.0.5 it runs with its first detector, which says so plainly when you're talking to the box twice as much as you were, and the other 27 and the cold-read comparison are planned for shade. I've committed to not calling LocalGhost finished without them.
Encryption, duress and backup // ghost.secd · ghost.mistd · ghost.watchd
ghost.secd is the only daemon the phone talks to. The archive sits on a LUKS2 volume whose key is sealed to the box's TPM, or wrapped by the PIN with Argon2id on a machine without one, and the phone's PIN unlocks it over mutual TLS with a certificate the box issued at enrolment. A lost phone can be retired by its key, from the box or from another phone, and from then on every call it makes is answered as if the box were down. Wrong PINs slow down and then lock out, and a wipe PIN destroys the key while looking like an ordinary unlock (The Honeypot Under Your Desk). A decoy volume behind a PIN of its own is planned for a later wisp, and more PINs with volumes of their own, with the two FIDO2 hardware keys for setup and recovery, for shade (1.x) at the earliest. ghost.mistd, which will encrypt, shard and spread your backups across other LocalGhost boxes in The Mist, is planned for poltergeist. ghost.watchd starts the rest of the fleet, restarts what falls over and is the only daemon that can read everyone else's metrics.
The hardware
Today LocalGhost runs on a Debian 13 machine you already own, with a TPM and a spare disk for the encrypted volume, and much faster with an NVIDIA card (Set Up a Box has the list). Once hardware ships it will also run on a box we build. The reference build as of June 2026 is a tower around a 12GB consumer GPU, about £1,130 in parts, and the development box runs Gemma 4 12B on a 12GB card at about 53 tokens per second (The Expensive Part Was Never the AI). Hardware ships after the software is stable.
The Local-First Directory
Outside the box, LocalGhost keeps the Local-First Directory, a hand-checked list of AI tools that run locally and keep your data on your own hardware. Projects can host a small freehold beacon file at /.well-known/freehold.json to pledge they run offline, export cleanly and have no kill switch. The crawler that validates beacons is still being written.
What makes LocalGhost different
You pay once, or nothing
ChatGPT Plus, Claude Pro and Google AI Pro each cost close to £20 a month in the UK, with power tiers around £80 to £160 and Google's top tier near £200 (June 2026 prices).[1] LocalGhost has no subscription, the reference tower is about £1,130 once plus electricity, and you can build your own from the open code and pay us nothing. For a heavy or multi-tool user the hardware pays for itself inside a year, and for a single subscription in two to three.
No company holds your data, including us
ChatGPT, Gemini and Claude keep their memory of you on the provider's servers, under terms the provider can change. LocalGhost has no servers that hold your data, no accounts and no telemetry, so there is nothing for us to sell, lose in a breach or hand over when a warrant arrives. The box asks the internet only for public things (prices, news, map data and its own updates), never with anything of yours in the request. The phone, when you ask something the box can't answer from what it holds, sends the question's words to the search engine, and a who-is question to Wikipedia on a box without its own copy. The weather where you are comes from a forecast the box pulls once a day for the world's three thousand larger places (since wisp 0.0.3), and the country the phone is in from the box's own map (since 0.0.4), so neither a weather service nor a geocoder learns where you are. The privacy page lists each request. What the model learns about you stays on your disk in open formats, which is the export button I went looking for when I left ChatGPT and couldn't find (The Model Trap).
It keeps working if we disappear
LocalGhost is MIT-licensed, runs on standard parts, and has no licence server, auth service or remote kill switch. If the company folds, your box keeps running the version you installed, a copy of the signed setup mirror on your own disk still works, and anyone can fork the code. The Mist, planned for poltergeist, is designed to need no central server either. Compare Limitless (formerly Rewind), which joined Meta in December 2025, switched off screen and audio capture in its Rewind app on 19 December 2025, and deleted the accounts of its EU and UK users.[2]
It's built to disagree with you
Jain and colleagues (2025) found that assistants grow more agreeable the more they remember about a user, and in their tests Gemini 2.5 Pro's agreement sycophancy rose 45% with memory switched on.[3] As far as I know no cloud assistant ships a component whose job is to argue with you, because disagreement looks like a bad session in an engagement metric. ghost.shadowd is that component, designed around 28 detectors and a cold-read arbiter on a separate, swappable model, and its first detector has run since wisp 0.0.1, with the rest planned for shade (How to Prevent Building Your Own Dictator Brain).
Duress mode for the bored border agent
Most disk encryption assumes nobody will ever make you unlock the device. In wisp 0.0.5 ghost.secd has a main PIN and a wipe PIN, and the wipe destroys the key while the phone shows the same unlock it always shows, so whoever is watching sees nothing unusual. A decoy volume is planned for a later wisp, and shade goes further, as many PINs as you want, each opening its own volume or wiping the box, with no metadata showing how many exist, whereas VeraCrypt is stuck with a fixed two-volume design that Schafer and colleagues showed in 2017 leaves detectable traces.[4] The threat model is the bored border agent and the person with a wrench, not a nation-state forensic lab.
Who uses LocalGhost
Nobody outside the test builds yet, since the first release only came out on 2 October 2026 and no hardware is on sale. The first three hand-built test units are going to people who help spread the word through the giveaway. These are the people it's being built for.
- Individuals who want a personal AI with long-term memory and don't want a company holding that memory. LocalGhost is built for one person's life, not for governments or corporations.
- Heavy AI users paying for two or more of ChatGPT, Claude and Gemini. At £100 or more a month, a one-off £1,130 box pays for itself inside a year, and at £40 a month in about two and a half.
- NAS and homelab owners who already run their own services on a UGREEN, Synology or Linux box and have, or can add, a 12GB GPU.
- People who track their own finances, health and habits and want it all in one place they can query, in open formats they can take with them.
- People who travel with sensitive data and want a duress option at the border that doesn't depend on a company's policy.
- Self-hosters and Go developers who want to write tallyd parsers, test the security model, or list a local-first project in the Directory.
Local is not for everyone, and it doesn't need to be. If the cloud's convenience is worth more to you than who holds the copy, that's a fair call.
The team behind LocalGhost
Vlad Cealicu, founder
I was born in Romania in 1986, grew up on a military compound where my parents worked for the air force, moved to London in 2009 and started my software career in 2010. I co-founded CryptoCompare (later CCData) and spent twelve years building it as CTO, growing it to about seventy people and leading the work on CCCAGG, a crypto price index built from roughly three hundred exchanges and regulated under the UK Benchmarks Regulation (An Index Is Not a Person).[5] CoinDesk acquired CCData in October 2024, I spent a year on the employee side, and I started LocalGhost in December 2025.[6]
How LocalGhost started
I wrote the manifesto, the site and the first architecture over Christmas 2025 because I thought the window for local-first AI would close before I felt ready (Inflection). The idea goes back further, to a CryptoCompare portfolio tool that had over a million monthly users in 2017 and that I kept deliberately limited, because making it useful meant a company holding everyone's financial life. Then in March 2026 I switched from ChatGPT to Claude and lost two years of personal context that had no export button, which confirmed the rest.
Who is on the team
Right now it's me and AI tooling (Claude for most of the coding), and I'm not pretending that's a stable position. My wife Cristina asks most of the questions that end up as posts, from what happens if someone steals the box to whether you can record other people and still call it private software. Paul, who I worked with for over eight years at CryptoCompare, reviews drafts and taught me what a good question looks like (How to Build a Critic Worth Listening To). If you want to help, the Build page lists what's needed.
Where to find us
How LocalGhost works
Talking to us
Email info@localghost.ai. You'll be talking to me, there's no ticketing system and no auto-reply, and I don't promise a response time. Bugs and code discussion belong on GitHub, and public conversation happens on X and LinkedIn.
Setting up a box
On your own machine, Set Up a Box has every step in order. Install the signed app on your phone, run the guided setup on the box, scan the QR it draws (scanning is the enrolment, the phone gets its own certificate from the box) and unlock with your PIN. There's no account, no onboarding questionnaire and no personality test (Day One explains why). Everything setup downloads comes from the signed setup mirror on this site, and if the mirror can't be reached, setup stops rather than fall back to anywhere else. A box we build will come set up the same way.
How long before it's useful
The ghost reads the traces already on your devices, so the aim is for it to be useful on the first day rather than after a multi-week ramp. Overnight memory consolidation of a heavy day takes about two hours on a 12GB GPU against more than a day on CPU, which is why the reference build has a GPU.
When you can get one
Each release line is named after a ghost, smallest first. Every 0.x release is a wisp, 1.x will be shade, and specter, phantom and poltergeist come after, and the roadmap lists what each one brings. The first five wisps, 0.0.1 to 0.0.5, came out between 2 and 5 October 2026, so you can set it up yourself on a machine of your own today and put the signed app on your phone. The box we build is planned for specter, after the software is stable. There's no date for that yet, and I'd rather not guess one. There's no mailing list either, so watching the repo is the most reliable way to see when something lands.
Key facts
- Company Name
- LocalGhost.ai Ltd, trading as LocalGhost, a private limited company registered in England and Wales, company number 17213100, incorporated on 12 May 2026
- Type
- Open-source, local-first personal AI system (software and hardware)
- Status
- On wisp, the first release line of the public roadmap. The current release is wisp 0.0.5, the server and a signed Android app, the fifth cut of wisp on 5 October 2026, the core daemons run every day on the development box, and no hardware is on sale yet.
- Founded
- December 2025 (the project), 12 May 2026 (the company)
- Founder
- Vlad Cealicu, former co-founder and CTO of CryptoCompare / CCData
- Team Size
- One founder, building with AI tooling
- Headquarters
- London, United Kingdom
- Website
- www.localghost.ai
- Core Offering
- A fleet of local AI daemons that builds a private, long-term memory from your text, images, audio and personal data, running entirely on hardware you own with no cloud dependency
- Pricing
- No subscription, ever. The core software is free and open source. Pre-built boxes will be a one-time purchase at parts and assembly plus a 30% margin, price not yet set (the June 2026 reference build is about £1,130 in parts). Future optional daemons may be sold as one-time packages, and what you already have keeps working if you don't buy them.
- Contract Terms
- None. No subscription, no account, no licence server. What you buy keeps working whether or not you buy anything else.
- Services
- Photo, video, trail, note, voice and health ingestion, a personal memory layer, prices and personal metrics, situational recall, an adversarial critic daemon (first detector), an encrypted volume with a TPM-sealed key and a wipe PIN, a signed setup mirror and the Local-First Directory. Planned are a decoy volume (a later wisp), more duress PINs and hardware-key recovery (shade), and peer-to-peer encrypted backup, The Mist (poltergeist)
- Licence
- MIT, source at github.com/LocalGhostDao/localghost
- Communication
- Email (info@localghost.ai), GitHub, X and LinkedIn. Answered by the founder, no ticketing, no guaranteed response time.
- Notable Clients
- None yet.
- Customers Served
- No paying customers. Three hand-built test units offered through the June 2026 giveaway.
- Projects Delivered
- Public architecture and daemon specs, 17 Hard Truths essays, the Local-First Directory and the freehold-v1 beacon schema. No software release yet.
- Competitors
- Cloud assistants with memory (ChatGPT, Google Gemini, Claude), capture and memory tools (Limitless, Screenpipe, mem0), and self-hosted AI (Khoj, Open WebUI)
- Social
- X @localghostai, LinkedIn, GitHub, founder X @vladcealicu
Frequently asked questions
What is LocalGhost?
LocalGhost is an open-source personal AI system that runs on hardware you own and builds a long-term memory from your notes, photos, audio and data. Nothing of yours leaves the box, it asks the internet only for public things like prices, news and map data, and there's no cloud account.
Can I buy a LocalGhost box today?
Not yet. LocalGhost is on its first release line, wisp, whose first five cuts (0.0.1 to 0.0.5) came out between 2 and 5 October 2026, and the box we build is planned for specter, after the software is stable. Three hand-built test units are going out through the giveaway. You can set one up yourself today on your own Debian machine, and Set Up a Box has every step in order.
How much will LocalGhost cost?
The core software is free and MIT-licensed. Pre-built boxes will be a one-time purchase at parts and assembly plus a 30% margin, and the price isn't set yet (the June 2026 reference build is about £1,130 in parts). Future optional daemons may be sold as one-time packages, and there is no subscription.
Does LocalGhost need an internet connection?
Not to work. Inference, storage and memory all run on the box, and setup can come from a copy of the signed mirror on a USB disk. With a connection the box fetches public things (prices, news when the phone is away, its own updates), and you reach it from outside through your own domain and port 443 with the phone's certificate. The Mist backups are planned for poltergeist.
Which AI models does LocalGhost run?
Open-weight models through llama.cpp on your own hardware. The box runs Gemma 4 12B with its vision projector for captions and chat, EmbeddingGemma 300M for search and Whisper large-v3-turbo for voice notes, and the phone carries Gemma 4 E2B for when the box is away. The model is a swappable dependency, a config change in ghost.oracled, so you can move to Llama, Qwen or Mistral.
How is LocalGhost different from ChatGPT memory?
ChatGPT's memory of you lives on OpenAI's servers, and the understanding it builds can't be taken to another model. LocalGhost's memory lives on your disk in open formats, and a separate daemon, ghost.shadowd, is there to disagree with you rather than keep you engaged.
What happens to my data if LocalGhost shuts down?
Nothing happens to it. Your box keeps running the version you installed, the code stays on GitHub under MIT, a copy of the signed mirror on your own disk still sets up a box, and nothing on the box calls home to us.
Why daemons and not agents?
A daemon owns the control flow. It runs predefined steps in a fixed order and calls a local model the way it would call Postgres, so the model never decides what happens next. An agent is the other way round, the model chooses the next step and the code follows it. LocalGhost is built so you never have to trust a model with that choice.
What if I lose my hardware keys?
In wisp 0.0.5 there are no hardware keys yet. The volume key is sealed to the box's TPM, or wrapped by your PIN on a machine without one, and there's no recovery path, so a dead board or a forgotten PIN means the archive on that disk is gone. Keep the originals of anything you can't lose somewhere else as well. Two FIDO2 keys, one with the box and one kept elsewhere that can restore your data to a new box, are planned for shade (1.x) at the earliest.
Who is behind LocalGhost?
LocalGhost.ai Ltd (company number 17213100, registered in England and Wales), founded and run by Vlad Cealicu, a London-based engineer who co-founded CryptoCompare / CCData and was its CTO for twelve years. He builds LocalGhost in the open with AI tooling, and you can reach him at info@localghost.ai or @vladcealicu on X.
How can I help?
Star or watch the repo, try wisp 0.0.5 on a machine of your own and report what breaks, try to break the security model (SECURITY.md says how to report it), or write tallyd parsers for the exports you use. The Build page lists the gaps, and donations in ETH go to zerocool.eth.
References
Last updated , when wisp 0.0.5 came out. The pricing row is the one I expect to rewrite next, when hardware goes on sale. If something here is wrong or out of date, email me and I'll fix it.