Features

What it does, how, and where it stops

Each section follows the same order: what it does, how it works inside, and its limits. Risk levels run from L0 (only reads) to L4 (runs code); the amber ones ask you first.

Commands and skills

Every BigOne capability is a skill: a skill.toml manifest, a main.py with exactly one Skill class, and test_skill.py. The manifest holds the skill's name, description, risk level, permissions, the JSON schema of its parameters and the regular expressions for the fast path. The description tells the model when to pick the tool.

Sourcesbuilt-in › user › generated; a user or generated skill can never take a built-in command away
Permissionsfs.read · fs.write · fs.delete · process.launch · process.kill · network · clipboard · code.exec
Disabled skillits code is never imported

LimitSkills you add yourself (user) are trusted code and are not checked. A skill with overly broad or invalid triggers is not loaded.

Skill	Risk	Fast path	What it does
list_folder	L0	yes	Lists what is in a folder
find_files	L0	yes	Finds files by name, extension, age
read_file	L0	—	Reads a code or text file with line numbers
recall	L0	yes	Searches long-term memory
list_grants	L0	yes	Lists folders BigOne may access
list_tasks	L0	yes	Lists background tasks
list_watches	L0	yes	Lists reminders and watches
open_folder	L1	yes	Opens a folder in File Explorer
open_url	L1	yes	Opens a website in the default browser
open_app	L1	yes	Starts an installed app from the Start menu
remember	L1	yes	Saves a fact or preference to memory
remind_at	L1	—	Reminds you at a given time
notify_when_done	L1	—	Tells you when a program ends
start_task	L1	—	Starts a background task it works on alone
cancel_task	L1	—	Cancels a task
cancel_watch	L1	—	Cancels a reminder or watch
revoke_access	L1	yes	Takes back a folder grant
create_folder	L2	yes	Creates a folder (reversible)
organize_folder	L2	yes	Tidies a folder by type or month (reversible)
edit_file	L2	—	Replaces an exact snippet in a file (reversible)
write_file	L2	—	Writes or rewrites a text file (reversible)
undo_last	L2	yes	Undoes the last reversible file operation
forget	L2	yes	Deletes memories on a topic; shows what first
draft_skill	L2	—	Drafts a new skill; runs no code
delete_files	L3	yes	Moves filtered files to the Recycle Bin
delete_folder	L3	yes	Moves a folder to the Recycle Bin
grant_access	L3	yes	Grants lasting read or write access to a folder
install_skill	L4	—	Installs a pending skill draft (15-min approval)

Files and "undo"

Organizing moves files into subfolders by type or by month; it never overwrites an existing file and makes "name (2).ext" instead. A snapshot is taken before editing or writing a file. Every reversible job is recorded; "geri al" (undo) reverts the last one.

Deleteto the Recycle Bin only; there is no permanent-delete code
Bulk jobsat least one filter, at most 2000 files per run
Readingat most 2 MB, no binary files; 12,000 characters per call
Neverdrive roots, the home folder, Windows and Program Files; secret files such as .env or SSH keys are never read

LimitFiles over 1 MB are not edited. Known folders (Desktop, Documents, Downloads…) cannot be deleted.

> İndirilenler klasörünü türe göre düzenle   # organize Downloads
  Onay gerekiyor (L2): 43 dosya taşınacak (Belgeler 23, Resimler 14, Arşivler 6).
  … Geri almak için: "geri al".
> geri al                                    # undo
  Onay gerekiyor (L2): Geri alınacak: … 43 dosya eski yerine dönecek.
  Geri alındı: 43 dosya eski yerine taşındı.

Long-term memory

What you ask it to remember is written to SQLite and turned into a 768-dimension vector locally, with embeddinggemma in Ollama. Each LLM request gets the 3 records closest in meaning to your message (cosine ≥ 0.45). Without the embedding model, search falls back to word roots and the missing vectors are filled in later.

Commands"hatırla" (remember) · "ne biliyorsun" (what do you know) · "bunu unut" (forget this)
Turkish testthe right record on top in 4 of 4 queries; correct ones 0.56–0.81, wrong ones ≤ 0.47
Retentionlong-term memory is never deleted automatically; chat is kept 90 days

LimitSearch runs over all records in Python; it is built for personal scale.

> hatırla: proje toplantısı her salı 10:00   # remember: …
  fast path · remember · L1
  Tamam, hatırlayacağım: proje toplantısı her salı 10:00
> toplantı hangi gündü?                      # which day was it?
  LLM · the 3 closest memories go into the system prompt

Voice

Speech is processed entirely on your PC. The wake word is looked for in the text Whisper produces, not in the audio signal, so no separate wake-word model is needed. The desktop app also has a Talk button: nothing is transcribed unless you press it.

Speedsmall on CPU ~3 s per sentence; turbo on NVIDIA also ~3 s, far more accurate; "BigOne, what's on my desktop?" answered in 3.7 s
Accuracyturbo on NVIDIA: word error rate 32.7% → 3.8%; wake word 8/8, 0 false alarms
Voice approvalup to L2, "evet / hayır" (yes / no); a negation always wins

LimitFast speech can garble Turkish suffixes. Deleting and running code cannot be approved by voice. Turkish only.

07Voice pipeline. Amber dashed line: the mic is closed while BigOne speaks, so it does not hear itself.

Companion

When you say "let's look at this", a work session opens in a folder you granted. The folder is scanned every 2 seconds, changes are batched and the strong model takes one look: mostly it stays quiet, if needed it suggests or warns, and in a write-granted folder it can make a reversible fix. It learns from what you say and how you correct it; background tasks and reminders run here too.

Restraintnever speaks before you have been quiet 20 s; at most 3 suggestions per 10 minutes; "be quiet now" lets only urgent warnings through
Learninga note starts at 0.4 confidence, +0.15 each time it repeats; passwords, health data and notes that grant permissions are never stored; it is told not to keep notes about other people, and in one eval it still did
Tasksstep by step with the usual approvals; asks after 5 steps or 15 minutes; after a restart it never silently resumes

LimitSpeaking up on its own and learning only run with the strong (cloud) model; the local 3B model was weak at this (see the measurements on the AI page).

08The mind layer. Amber: the attention part that looks at file changes.

Self-extension

When you ask for something no skill can do, BigOne offers to write a new skill. The model writes the manifest, code and test; a static check inspects allowed imports, forbidden constructs and risk consistency without running anything. To install it you see the full code of all three files on screen and approve; only then do the tests run, in a separate sandbox process.

Sandboxtemp copy, clean environment, Windows Job object: one process, 1 GiB memory, 60 s
Versionsan installed skill is committed to git; rollback or removal is one command
Corenever changed; only new skill folders are added

LimitIn-process isolation can be bypassed; the real boundary is you reading and approving the code. OS-level isolation (AppContainer) is deferred.

09The path of a new skill. Amber: the step where you see and approve the code.

Phone

The Android app is a remote control for BigOne on your PC: text chat, push-to-talk and approval notifications. Audio is recorded on the phone and transcribed on the PC with Whisper. Pairing works with a QR code or, without a camera, by typing it in; both screens show the same 16-character verification code.

Phone tokenchat, voice commands, approvals, event stream; cannot change keys, permissions, devices or voice settings
Approvalup to L2; deleting and installing code only on the PC screen
TechKotlin, Jetpack Compose, OkHttp, ZXing (no Google Play Services), Android Keystore

LimitSame Wi-Fi or VPN only; no port forwarding. No fallback intelligence on the phone: if the PC is off, the app says so. Android 8.0 or newer.

Phone tab: QR code, 8-character pairing code, PC address, port 8766 and the verification code — 10"Add phone" on the desktop: the QR, the details for typing it in, and the paired phones.

Desktop app

A Tauri app that lives in the system tray. If the core is not running it starts it without a window, and it only stops a core it started itself. Tabs: Chat, Tasks, Notes, Grants, Watches, Phone, Models, Trust. The approval window always opens on top; if the window is hidden you get a Windows notification.

Tray ringblue: here · amber: working with you · grey: away or unreachable
Modelsturn providers on/off, model, key (write-only), order, "Test"; no restart needed
Bridgethe Rust bridge only allows /v1/ paths; the token never reaches the web UI

LimitThe installer is unsigned: SmartScreen warns about an unknown publisher. No in-app updates; a new installer goes over the old one and keeps your data.

Models tab: Gemini with a quota-exhausted badge, Groq without a key, Cloudflare needing an Account ID — 11The Models tab: free providers, when Gemini's exhausted quota reopens, Groq without a key.

Read on

Security →

Approvals, grants, earned trust, the phone link.

AI →

Which model when, measured results.

Technology →

Stack, decisions, packaging, tests.