Shadow Shell AI is a private, uncensored AI service with something the privacy-first chat apps don’t have: hands. It doesn’t just talk about the job — it runs the commands, reads the output, fixes its own mistakes, reaches the machines you point it at, and comes back with the result — words, code, or the picture the job called for.
Sign in and start. No filter deciding which questions you’re allowed to ask, no company mining your conversations, and a vault that encrypts your history with a passphrase we do not hold and cannot recover. And when a job needs more horsepower than one model has, one toggle lets Shadow quietly consult a frontier model and fold the answer in — or summon a whole council of them.
the nightly export on build-2 failed again. find out why and fix it.
reading /var/log/export/nightly.log · 412 lines · last run 03:10
cause: /srv/export is 97% full — 38 archives older than 30 days
[review] rm /srv/export/*-2026-0[6-8]*.tar.zst — PASS · live tree untouched
freed 41 GB · re-ran the export · 2m 14s · checksum verified
done. want a weekly prune added to the cron so it stays fixed?
The big labs ship genuinely capable agents — Claude Code, Codex, agent modes that run commands and edit files — but behind a content filter that decides which questions are permitted, on servers where the company can read everything you type. The privacy-first apps took the filter and the reading away, and kept the text box. Shadow Shell AI keeps the agent, drops the filter, and makes your data unreadable to us.
Built on Shadow v4.8, our open, uncensored model. No refusals, no lectures, no topics quietly walled off. It works for you, not for a policy team.
Real tools: run commands, read and write files, execute code, connect out to the machines you point it at. Give it a task, get an outcome — not a suggestion.
Your history and your files are encrypted with a passphrase only you know. What sits on the server is ciphertext — unreadable to us, to staff, and to anyone who seizes it.
This is not a roadmap. Each item below is running in production and in daily use.
It plans, runs real commands, reads what came back, corrects itself and reports what actually happened — the difference between an assistant that drafts instructions and one that finishes the task, with every step on the record.
It can fetch pages, call APIs, clone a repository and talk to the servers you give it. Every task gets its own clean workspace, and the network path it uses is governed and rate-limited rather than wide open — reach without a blank cheque.
Feed it whole codebases, long documents or a sprawling multi-day thread. Long threads are summarised as they grow, under a retention contract: facts you pin are never dropped, and it can search the whole conversation verbatim on demand.
Drop in documents, code, logs, spreadsheets, PDFs or images. They land in the task’s own workspace, where the AI can open them, run them and take them apart rather than skim a preview. Screenshots, scans and photos are read, described and transcribed.
Ask for a picture in any house-model conversation and it generates straight into the chat’s Files drawer — anime or photoreal, no mode to switch. Titles are set in real fonts with the spelling checked, and flat text edits are free. A section of its own →
Queue your next message while the current one is still running and it starts the moment a lane frees. Press stop and it stops — immediately, cleanly, with the partial work still on screen, so you can redirect rather than wait it out.
Long jobs survive dropped connections, flaky networks and closed laptops. A chat can also wake itself: pick a time, or let the AI pick one, and it comes back on its own to check the download, finish the build and carry on — finished, not abandoned.
On genuinely hard problems Shadow can open a discreet consult with a frontier model — Claude, GPT, GLM, DeepSeek — on your own API key. It writes the question itself, folds the answer in and keeps working as one conversation. Its own section →
A distraction-free terminal aesthetic that works the same on a desktop and a phone. Conversations in a tree, notes alongside them, syntax highlighting, keyboard shortcuts, and answers that stream live and reconnect on their own when the network blinks.
Seats, plans and per-person limits, plus 30 days of real usage for every person and optional automatic limits that step in before one runaway task spoils the day for everybody else. One bill, one console, and numbers you can trust.
Every new chat opens on a compact row of modes — Code & Ops, Research, Writing, Marketing, Health, Legal and more — each with its own approach, starter prompts and an optional profile it remembers for you, sealed under your vault key. See the modes →
Set a 4- or 6-digit PIN on a phone or laptop you trust and unlock the vault with it instead of the full passphrase. The PIN never leaves the device, works only on that device, and is wiped after five wrong tries; the passphrase is always the way back in.
The market split in two. The big assistants have excellent agent tooling — Claude Code, Codex, agent modes that run real commands — paired with a policy filter and a provider who keeps the transcript. The privacy-first chat apps fixed the surveillance and stayed a text box — no tools, no ability to act, often a history stranded in one browser. Shadow Shell AI is the overlap: the agent, without the filter or the log.
| Mainstream AI assistants | Privacy-first chat apps | Shadow Shell AI | |
|---|---|---|---|
| Answers without a policy filter | ✕ Refuses by design | ✓ | ✓ |
| Real tools — can do the work, not just describe it | ✓ Claude Code, Codex, agent modes | ✕ Text only | ✓ Full agent tooling |
| Operator cannot read your conversations | ✕ Stored and reviewable | ~ Varies; often browser-only | ✓ Zero-knowledge vault |
| Attach files the AI can actually open and run | ✓ Upload, run, analyse | ✕ | ✓ Into the workspace |
| History follows you across devices | ✓ | ~ Often trapped locally | ✓ Encrypted, synced |
| Consult a frontier model on demand | ~ That vendor’s models only | ✕ | ✓ Consult Mode — any provider, your key |
| Generate images without a content filter | ~ Filtered and logged | ✕ Rarely offered | ✓ Built in — titles rendered in, spelled right |
| Specialist modes that remember your setup | ~ Custom instructions, readable by the provider | ✕ | ✓ 11 modes, profiles sealed under your key |
| Context window | ✓ Large; varies by tier | ~ Typically modest | 256K tokens |
| Memory inside a long thread | ~ Summarised past the window; saved memories, where offered, sit readable on their servers | ~ Short threads; nothing to pin | ✓ Retention contract, pinned facts, searchable full history — sealed under your key |
House-model conversations now generate images natively. There is no mode to switch, no separate app, no different client: the model recognizes when the job needs a visual and fills the request itself, right in the thread where you were already working. The result lands in the chat’s Files drawer beside your code and documents — newest first, downloadable one or all — and the conversation simply carries on. Uncensored, encrypted like everything else, and bounded by honest published numbers instead of mystery.
The same philosophy that governs the chat, now for pictures: no banned-prompt list, no classifier deciding which images you are allowed to make, and what you ask for is what gets drawn. Adults making art for adults is your business, not ours. The ordinary limits of the law still apply — drawn by the Terms of Service, not by a keyword — and the one hard line, sexual content involving minors, is blocked everywhere, always.
Two purpose-picked engines: anime for illustration, photo for everything else — realistic people, posters, product shots, painting — both driven in plain language, no tag dialect to learn. It edits too: point it at any picture in the chat and say what to change — make it night, swap the background, add a hat — and the composition stays. Eight sizes from square to wide banner to phone story, up to two images per request.
Photo-style images letter themselves: quote the words and they are painted into the scene — a shop sign, a poster headline, a neon tube. On anime they are set in a real font — 53 typefaces from poster to gothic — and rendered into the scene as a material: neon, chalk, carved wood. A vision model reads the result back and checks the spelling; a miss gets one retry, then the flat version or a plain note of what it read — never a silent typo. Flat titles stay free: no GPU time to change the words.
On a vaulted account, generated images are sealed under your own vault key whenever no reply is being written in that chat — file names included — and sit on our disks as ciphertext. They are opened in the chat’s walled-off workspace only while a reply is in progress, and sealed again the moment it finishes. Generation is logged by prompt hash, never by raw prompt, and nothing you generate is used for training.
working · 2 images · filed to the drawer
[image] 832×1216 + 1216×832 · anime · in the Files drawer
you: title it “NIGHT SHIFT” — pink neon tubes, Monoton
[image] title rendered into the scene · spelling verified
you: flat version too — Bebas, #39FF9E, on the lockup
[image] re-composited · instant · zero GPU spent
Every model, sooner or later, meets a problem at the edge of what it can reason through alone. Most services leave you stuck right there. Shadow Shell doesn’t: flip one toggle per conversation and your Shadow quietly opens a consult with a super-frontier model — on your own API key — gets the help it needs for the task at hand, and carries on as one seamless answer.
Your Shadow composes the consult itself: a focused, professional question carrying only the context it chooses — never your raw conversation tipped over the wall. Work that would stall in a hail of template refusals anywhere else gets a clean, direct technical answer here.
Claude, GPT, GLM, DeepSeek — plug in the API keys you already have and Shadow consults them on demand, at your provider, on your rates. Keys are sealed inside your vault and never touch the sandbox your AI runs in.
It doesn’t stop at a second opinion. Run the whole agent as one frontier model — full tools, full workspace — and let it consult a third. Or summon a council: every provider you’ve keyed answers the same question blind, in parallel, and Shadow states where it lands. The platform section below has it →
A consult is a real egress: one question plus chosen context, to the provider you picked. So each one is written into the conversation where you can read it, and any chat that ever used one is marked permanently — even after the toggle goes back off.
planning · 4 steps · consult ON
[consult] claude — “framing pattern for length-prefixed binary over TLS?”
[consult] answered · 11 lines · folded into step 2
done — workspace/client.js, two edge cases the consult caught
A new chat opens on eleven modes. Each one sets how Shadow works — its approach, the tools it reaches for, the questions it asks first — and you can switch at any time from the composer. Eight are for work. Three are personal: they give information, not advice, and ask you to confirm you are 18 or over the first time you open them.
General, Code & Ops, Research, Writing, Tutor, Business, Marketing and Create. Every one of them is the full agent — shell, files, the wider internet and Consult Mode — with its own approach and its own first questions, and Create opens with the image studio built in. Each mode ships with starter prompts, so a blank chat is never a blank page, and you can switch modes mid-conversation from the composer.
Health looks things up in public drug and health references and cites them. Legal searches case law, and every answer gets a citation check underneath — each case marked found, mismatched or not found. Wellbeing is a calm place to think out loud; it is an AI, not a person and not therapy, and it says so. These three run no shell commands, and Wellbeing never contacts an outside model.
Give a mode the background it keeps asking for: your stack and environment in Code & Ops, your voice and audience in Writing, your brand, offer and channels in Marketing, your medications and allergies in Health. A profile is used only in its own mode, can be changed or cleared any time, and is sealed under your vault key like everything else — lock the vault and nobody, us included, can read it.
If a message reads like someone in crisis, in any mode, a helpline card for your country comes first and the answer is checked before it is shown. Requests for a method or a dose of self-harm are not answered. Sexual content involving minors is blocked everywhere, always — chats, image prompts and profiles. The checks are automatic; no person reads your messages to run them.
An agent that runs commands and spends money needs discipline built in, not bolted on. The pattern holds across everything below: toggle-gated per chat, budgeted per turn, carded in the transcript, and it fails in the safe direction. You get the leverage; the machine keeps the receipts.
Summon a council and every provider you’ve keyed — in-house Shadow included — answers the same question blind, in parallel, one card each. No model sees another’s answer. Your agent reads them all, weighs them, and states where it lands and why. Second opinions are cheap; groupthink is what’s expensive.
Opt in per chat and an external model vets destructive shell commands before they run. Four reviews a turn, verdicts cached, and two consecutive denials lock the turn — fail-closed. If the reviewer is unreachable the command still runs, under a loud unvetted card, and every concern is shown verbatim.
Spawn a subagent on our own GPU pod and the bulk reading burns the child’s context, not yours — the parent grows only by the capped result. Each child gets its own scratch directory, inherits the egress gate and the review hook, and comes back as one collapsed line. Two per turn, two concurrent, 240 seconds each.
Move a job to another chat in your account and the conversation compiles the complete brief itself — goal, state, paths, constraints. On the far side it waits as a card with one button; nothing runs until someone presses it, and both transcripts keep the record. Your account, your continuity.
working · review ON · 4 reviews / turn
[review] rm -rf /staging/current — DENY — “current is a live symlink into prod; the tree you want is /staging/v3”
[review] rm -rf /staging/v3/build — PASS · circuit 1/2
rebuilt — /staging/v3/build, the live tree untouched
Most “we don’t log” promises are policy — a company choosing not to look, which it can reverse and a court can override. Ours is architecture: what sits on our servers is mathematically unreadable to us.
Shadow Shell AI is a hosted service — you sign in and use it; the model, the tools and the capacity are ours to run. Seats are how it is sold: a light seat for yourself, a working one, a heavy one, or a set for a team.
The full 256K context window on every plan, same price.
No, and not as a promise — as arithmetic. Your history and your files are encrypted under a key derived from a passphrase that never leaves your control. We hold ciphertext. A subpoena gets the same ciphertext.
That history is gone. There is no reset link and no support ticket that recovers it. That is the cost of the guarantee being real, and we would rather say so on the pricing page than in an apology email.
There is no policy filter sitting between you and the model, and no topic list. What remains are the ordinary limits of any hosted service — you are still responsible for what you do with it, and abuse of the platform’s own resources is bounded. One floor never moves: a message that reads like a crisis gets a helpline card first and a checked answer, self-harm methods and doses are not given, and sexual content involving minors is blocked outright.
Its own per-task workspace, the files you attach, and the outside network. It cannot touch the host, other accounts, or anything on your own machines unless you deliberately give it the way in.
It carries on. Older turns are folded under a retention contract: the summary keeps your goals, every decision and correction, the exact values that matter and the state of the work, and anything it cannot carry verbatim is marked unverified instead of guessed. Facts you pin in Chat options are put back in front of the AI after every fold, and it can search the complete conversation — every message, command and result since turn one, sealed under your key — and quote it back exactly. Say “remember that…” in the chat and the fact is pinned for you.
One toggle per conversation. When it’s on, your Shadow may — at its own judgment — send a focused question to a frontier provider you configured, on your API key, and fold the answer back into its work. What leaves is the question and the context Shadow chose, not your whole transcript. Every consult is visible in the conversation, any chat that ever used one stays marked, and with the toggle off, nothing leaves at all.
Optionally, a second opinion. Review-before-destructive is a per-chat toggle: when it’s on, an external model vets destructive shell commands before they run and its concerns are shown verbatim. Two consecutive denials lock the turn outright. If the reviewer can’t be reached, the command proceeds — but the transcript gets a loud unvetted card, so you are never quietly trusting a service that was down.
Yes, with the same shape as the chat: no banned-prompt list and no classifier between you and the canvas, and the ordinary limits of the law still applying. The platform’s own resources are protected the honest way — published numbers, not mystery: twelve images per seat per hour, four per minute, with flat text edits unlimited and free. The GPU pod stays warm around the clock, so an image takes about thirty seconds and one with an in-scene title about a minute — the first of the day included.
You keep working. Past the line, the seat drops to a single lane at a reduced request rate until the daily reset at 00:00 UTC — requests wait their turn rather than fail. Only at three times the allowance, which is automation running around the clock rather than a person, does a seat pause until the reset. Nothing is deleted, nothing is billed as overage, and you can watch the number all day rather than discovering it at the end of the month.
Any month, both directions, yourself. On a PayPal plan your Client Area switches between Core and Operator and adds or removes Team seats in a couple of clicks: moving up and adding seats take effect at once, while moving down and removing seats take effect at your next bill, because the current month is already paid for. Paying by Bitcoin, or want anything else? Write to [email protected] and it is done on your account.
Yes — that is what the Health and Legal modes are for. They explain, look up public references and show their sources, and Legal checks its own citations. They are information, not advice: no doctor, lawyer or counsellor relationship is formed, and for anything that matters, take it to a professional. In an emergency, call your local emergency number.
Not on your own devices. Set a 4- or 6-digit PIN for each phone or computer you trust and unlock with that. The PIN works only on the device that set it, five wrong tries erase it, and the passphrase always works as the way back in.
To chat, no. To get the most out of the tools, it helps — this is a product for people who want the AI to finish the job, and finishing the job usually means touching real systems.
No. We could not if we wanted to.