→00VSCode Extension · Marketplace
The memory secretary
for your project.
viway watches your AI agent’s dialogue and, after every reply, extracts decisions, facts and open questions into a living knowledge base on its own. No manual work. Context is never lost — not even after /compact, a restart or a folder move.
→How the processes flow
Every subsystem is a clear flow you launch from the dialogue and watch in the dashboard.
The agent replies → auto-trigger → the secretary reads the turn → structured operations on .viway/ → a living wiki of 8 sections. Zero manual work on your side.
A task to the fella → it works in its own session beside you → a structured report → viway checks that report against the real files. It asked something — answer, and it carries on from the same place.
The plan is filed in the project memory → the agent pulls the next step and a reality check → does it → reports → the secretary reconciles the report against the project and amends the plan. The plan survives compaction and a change of session.
The page opens inside the editor → the agent reads it as a text outline, clicks and types → the human annotates the same page on top → the edits reach the agent as a queue. The console and failed requests survive a page reload.
The human switches «Hands» on and ticks which windows the agent may touch → the agent looks into the window, reads its elements as a text outline → clicks and types → every action lands in the journal. It cannot leave the ticked windows, and «Stop» cuts it off mid-sentence.
A prompt from the task → approval → your Codex or Grok, or an OpenRouter key paying per picture → the finished file into staging, into the project on request. The human’s marks on the canvas are read by the agent as instructions.
A prompt or an existing still → an order on an OpenRouter key paying per clip → the clip takes minutes and survives a window reload → viway collects the result itself and puts the file into the project.
A question from the dialogue → query / recall / semantic → the secretary searches .viway/ → a precise answer with citations. The memory answers right in the chat.
→01Key features
The secretary — a separate agent — maintains the knowledge base on its own after every reply: it classifies the turn, decides what matters and updates the right wiki section. Zero manual work — you just work.
/compact, a VSCode restart, a folder move — your project context stays put and is never lost.
viway collects dialogues from every source: parallel Claude Code, Codex CLI and GitHub Copilot sessions see one shared knowledge base — and do not redo work already done.
Four providers to choose from. One key — and you decide which model runs your memory.
The screen, mouse and keyboard of this machine, lent to the agent for a while. It looks into an application window, reads its elements as a cheap text outline, clicks and types. For work that is neither a page nor a file: an installer, a licence dialog, a CAD seat, a settings window. Only a human switches it on, every action is written down, and «Stop» cuts it off mid-sentence.
The agent queries the memory straight from the dialogue — “what did we decide about X?”, “have we seen this before?” — via query, recall and semantic search, and gets a precise answer with citations from the wiki.
Everything viway does lives in one isometric panel: Secretary, Fella, Planner, Communicator, Browser, Images, Video, Hands and Knowledge base. A real-time turn feed, colour-coded statuses, metrics and a spend audit — on the same screen.
Fonts are bundled, with no external CDN dependencies. The interface opens even with no network.
query · recall · catchup · note · planner · fella · browser · computer · images · video · playbook — the bundled skills install with one click.
A page opens right inside the editor — a file, a folder, a dev server or a live site with real cookies. The agent reads it as a cheap text outline instead of a screenshot, clicks and types, sees the console and the failed requests. And the human annotates that same page on top: comments, arrows, moves — and they reach the agent as a queue of edits.
A living plan watched over by the secretary: the agent authors the steps, the secretary reconciles every report against the reality of the project and amends the plan to match. The plan lives in the project memory rather than in a forgotten file — and survives context compaction, a change of session and closing the editor.
Icons, covers and mock-ups straight from the task — on your Codex or Grok subscription, or on an OpenRouter key paying per picture. The human annotates the canvas over the result, and the agent reads those marks as direct instructions. Into staging by default, into the project only on an explicit request.
Semantic search via embeddings finds what you need even when the words don’t match. And recall pulls ready solutions to past problems — the symptom → cause → fix triple.
Project playbooks: turn a working procedure into a skill the agent finds and runs on its own — with an optional enforcement gate for critical steps.
Before every run viway snapshots the knowledge base — roll any result back with one pick from the list.
A one-click reviser cross-checks the knowledge base against the real code + git: it finds closed-but-still-open questions, stale references, stack conflicts (wiki says Tauri while the code is Electron) and broken links in docs. Report-only — you apply the findings yourself, in bulk or one by one.
A second working session beside yours: a headless Codex or Claude on your own subscription — or viway’s own agent loop on an OpenRouter model. Hand it a task and get on with yours; it reports in structure, and viway checks that report against the real files. It asked something — answer, and it continues in the same session.
A short clip from text or from an existing still — on an OpenRouter key, paying per clip. The order takes minutes and survives a window reload: viway collects the result itself and puts the file into the project.
→02How it works
The secretary fires on its own after every agent reply. You just work — the knowledge base fills up in the background.
You work with Claude Code, Codex CLI or Copilot as usual — in VSCode, Antigravity or Cursor.
Every agent reply launches the secretary — no buttons, no commands.
The model analyses the dialogue and picks out what’s worth remembering.
Decisions, facts and questions become tidy records.
Everything lands in a structured knowledge base right inside the project.
→03Problem & solution
An AI agent is brilliant in the moment but forgetful over time. Every /compact, restart or new session wipes what you’d already decided. viway turns the stream of dialogue into durable project memory.
/compact and restarts→04Knowledge base
Project memory is organised like a personal archive — each section owns its own layer of knowledge. Tidy, readable, under control.
→05Multi-provider
Connect any of four providers. Switch the model to fit the task and budget — viway stays the same.
A single access point to dozens of models through one key.
Direct access to the Claude model family.
Qwen models via Alibaba Cloud.
The Gemini family for a flexible balance of price and quality.
→06More power under the hood
Hands
When the work is neither in a file nor on a page — an installer, a licence dialog, a CAD seat, a system settings window — the agent looks into that application window and acts inside it. Only you switch it on, you tick which windows it may touch, and you see every action in the journal.
Communicator
→07Savings
Request caching cuts repeat costs — up to ~75% savings. Memory works in the background and won’t break the bank.
The secretary reuses already-processed context instead of paying for it again. The longer a project lives, the bigger the savings.
→08Privacy & control
The knowledge base lives right inside the project. Before every write — an auto-snapshot, one-click rollback. And crucially: viway never touches your project code.
Memory is stored locally in the project folder. No cloud storage — your data stays with you.
An auto-snapshot before every write. Something off? One-click rollback restores the previous state.
The extension works only with its own knowledge base. Your source code stays untouched.
→09See it in action
On the left — settings and provider selection; on the right — a live turn feed with colour-coded statuses and metrics. Everything the project remembers, right in front of you.
→10Install
Find viway in the VSCode Marketplace or install it by the id viway.viway-v3. One button — and the extension is in your editor.
Pick one of the four providers and paste your key. OpenRouter is an easy start: one key unlocks dozens of models.
Turn on the bundled skills query · recall · catchup · note · planner · fella · browser · computer · images · video · playbook — and the agent can use the memory and every subsystem directly.
From there viway does everything itself. After every agent reply the secretary updates the living knowledge base — and you watch it on the dashboard in real time.
viway 0.75.0 is already on the VSCode Marketplace. Install it in a minute — and stop losing context.
→12Contacts
A question, idea, bug or suggestion — drop me a line. I answer personally.