AI Assistant · Step 1
Press Ctrl + I on any screen and a chat panel docks beside what you are looking at. Say what you need — "sign in to Gmail on these 50 profiles and export the inbox" — and the assistant builds the workflow node by node, test-runs it on a real profile, reads the log and keeps fixing until it passes. You review the result on the canvas, not the code.
Ctrl + Ior the ✨ icon in the title bar · sessions persist, run in parallel, and survive switching screens
01 · Edits the workflow itself
Full CRUD over workflows and every node inside them — create, rename, duplicate, wire edges, change config. Changes go through the editor's own validation and auto-save to the .hideauto file; undo still works.
02 · Runs, inspects, repeats
run_test on a live profile, read the per-step log, inspect the browser (URL, query, eval, screenshot), stop a run. Up to 24 tool steps per turn, so it iterates edit → test → fix without you nudging it.
03 · Knows what you see
Every turn carries a one-line context — which workflow is open — plus your profiles, variables and Critic findings on request. Paste a screenshot and vision-capable models read it.
04 · Your model, your key
OpenAI-compatible, Anthropic or Google endpoints, or sign in with your Claude account. Four fields: type, base URL, key, model. Keys live in the main process and are never shown or sent to the UI.
One request, the whole loop
Finds the target workflow by name, reads its graph, checks which node types are valid.
Applies a validated change set and saves. Every tool call shows as a chip you can expand.
Test-runs, reads the first error, inspects the page if needed, and edits again.
Runs a Critic review, summarizes what changed, and can commit it to git when you say so.
Bring your own model
Tool-calling quality depends on the model. The settings screen suggests models that handle it well.
Sandboxed by construction
Workflow · Step 2
Everything the assistant builds is an ordinary workflow you can open, read and edit. Eight node groups, a Critic review before anything ships, hundreds of profiles on a schedule — and nothing ever leaves your machine.
Drag a node from the palette and wire it through two output ports — success and error. Branch with IF / IF-ELSE / Element exists, repeat with Loop, halt with Stop. Variables travel through everything in double-brace syntax.
Node palette · 8 groups
Why a gate
A positional selector, an empty error port, a delay hand-set to zero — a single test run shows none of it. Three hundred scheduled profiles at 2am show all of it, and by then the data is already wrong.
resilience
Running once does not mean running twice.
anti-detect
Machine traces: cadence, ordering, dwell time.
data-safety
Losing data, or reporting success when it actually broke.
cost
Money and time: redundant calls, writes, runs.
How the score works
The score says where the flow stands. Shipping is blocked by the count of errors, not by the score — so a real bug never slips through because the rest is too clean.
Git & Backup · Step 3
Your workflows folder is a real git repository from the first launch — no git install, no setup. Commit when a change is worth keeping, read the diff as a graph, restore any version, and push to a remote you control when you want a copy off this machine.
Versioning without installing git
A pure-JS git engine ships inside the app. Init is automatic and idempotent; existing workflows get an initial commit. The repo contains only .hideauto files — no database, logs or tokens.
Commit on purpose, not on save
Save all you want; nothing is committed until you write a message. Stage per workflow or commit all, mark a milestone tag, and see a badge on the nav rail counting what is still uncommitted.
Diff as a graph, not as JSON
Side-by-side or overlay: which nodes were added, removed or changed, drawn on the canvas. Switch to text when you want the raw lines. Preview or restore any commit — restores land as uncommitted changes, history is never rewritten.
Backup to your own remote
Paste an HTTPS URL — GitHub, GitLab, Gitea, your own server. Test the connection, then Backup (push) and Pull. If git is on the machine, sign-in goes through your browser; otherwise a token stored encrypted with the OS.
From edit to backup
The workflow file changes on disk. The rail badge ticks up.
Message, optional milestone. Commit one workflow from the editor or everything from the Changes panel.
History per workflow, commit graph for the repo, graph or text diff for any pair.
Push to origin. Pull fast-forwards; if both sides changed, a merge screen lets you pick per workflow.
Two machines, one remote
The app is only a git client. It never hosts a remote, never uploads anything unless you configure a URL and press Backup.
Product figures
[placeholder — awaiting real figures]
nodes with a real handler
step ceiling per run
Critic rules, 4 groups
bytes of workflow sent to a server
Compare
Not because the code is bad, but because every change is an afternoon, nobody reviews it, nobody knows who changed what, and nothing tells you where run 217 stopped.
| Features | HideAuto | Hand-written script | Generic automation tool |
|---|---|---|---|
| AI assistant that edits, test-runs and fixes the flow | ● | copy-paste from a chatbot | partial |
| Visual building, no code required | ● | — | ● |
| Anti-detect profiles with their own fingerprint | ● | roll your own | — |
| A gate that blocks the build while the flow has errors | ● | — | — |
| Git commit, graph diff & backup to your own remote | ● | git, but the diff is text | — |
| Schedules + batches across profiles in parallel | ● | cron + DIY | partial |
| Per-step logs that pinpoint where it stopped | ● | console.log | partial |
| Package a flow as a .exe with sha256 | ● | — | — |
| Runs locally — data never leaves the machine | ● | ● | usually cloud |
Pricing
No workflow cap, no profile cap, no card. The AI assistant runs on your own model key. When real pricing lands, beta users hear it first — and everything you built is still a file on your own machine.
This beta includes: workflows · AI assistant · critic · schedules · profiles · version control & backup. The AI assistant uses your own model key or Claude login — HideAuto adds no markup.
Beta
/ month, during the beta
FAQ
Anything else, just ask — the beta moves fast.
Any OpenAI-compatible endpoint (OpenAI, OpenRouter, Ollama, vLLM…), Anthropic Claude via API key or your Claude subscription login, and Google Gemini. You bring the key; requests go straight from your machine to the provider. HideAuto adds no fee and never sees the key.
Blog
Kế hoạch ra mắt các tính năng đột phá: Tự động tìm phần tử thông minh bằng AI khi giao diện web thay đổi và chợ chia sẻ kịch bản workflow.
Phân tích ưu điểm vượt trội của mô hình Local-first trong HideAuto: không phụ thuộc Cloud, bảo vệ tài sản số tuyệt đối và giảm thiểu rủi ro rò rỉ dữ liệu.
Tại sao các hệ thống Cloudflare Turnstile, DataDome phát hiện được bot qua vi sai kết xuất font chữ và cách HideAuto spoofing chuẩn xác.
Free during the beta. Runs locally on Windows. Bring your own model key — no card, no cloud account.
Self-updating launcher · Bilingual VI / EN · Your data stays local