wae : local-first, AI Flight Recorder & Context Summarizer That Helps You Step Away

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass

Build locally. Capture automatically. Step away.

What I Built

wae* [work attention equilibrium] is a zero-config, lightweight, port-free, human-agent session anchor.
Securely captures the terminal footprints so you can stop watching the screen and trust the background execution.

You run your agent through it (wae claude, wae bash, any command), and wae quietly records a session, strips ANSI noise and secrets, then a detached worker asks a local Ollama AI model for a 3-bullet note and drops it into your vault/folder as 3-bullet note: what was done, what broke, and where things stand. The note lands as a plain Markdown file in any folder you choose.

How it gets you off the screen:
With the rise of dense, autonomous development agents like Claude Code, Aider, and custom multi-agent loops, developers are trapped watching walls of raw console noise, looping build spinners, and scrolling progress text streams, agent sessions pull you into babysitting a terminal because you're afraid you'll lose track of what happened.

• wae removes that reason. Start the session, go for a walk/touch grass or if got away from desk, and come back to three bullets instead of a thousand lines of scrollback. Everything runs on your machine, so there's nothing to keep checking. It let you trust, preserved context, and freedom to step away.

Who it's for: developers who run coding agents and automated shell loops and want a record they can trust without a cloud service watching their terminal.

Demo

Demo : https://drive.google.com/drive/folders/1n4CGs5A6A-NNa1FV_FCunqp3oCL_VNWY?usp=drive_link

wae --set-output ~/notes     # choose where summaries go (any folder)
wae bash                     # record a session
# ...work, then exit: the prompt returns instantly

A few seconds later the note appears:

# 🌿 wae session: `bash`
## Summary
-  ACTIONS: ...
-  ERRORS: ...
-  STATE: ...

Code

GitHub: https://github.com/divya-o/wae

How I Built It

Open-source AI at the core: local inference through Ollama with an open-weight model (Gemma by default, switchable with WAE_MODEL). The model only ever talks to 127.0.0.1.

• Capture: node-pty mirrors the session to your screen first, then buffers it. Keystrokes go straight to the pseudo-terminal with no middleware.
• Sanitizing: a regex filter removes ANSI graphics and redacts API keys and password= strings before anything touches disk.
• Storage: SQLite in WAL mode, so a crash or hard kill doesn't lose what was captured.
• Context protection: raw logs never go to the model. A line-by-line scanner keeps only installs, file changes, git actions and errors (max 50 lines), so small local models don't choke.
• Zero-lag handoff: the summary runs in a detached child_process.fork with .unref(), so the terminal returns instantly.
• Hygiene: raw logs are deleted only after the note is safely written.
• Optional durable mode (wae --durable): an open-source Temporal workflow keyed by the session ID retries the summary if Ollama is unavailable, and never writes the same note twice. If Temporal isn't installed or running, wae falls back to the standard summarizer.
• MCP server (wae --mcp): talks over stdio only (no ports), so an agent can read its own recent summaries.

The default mode opens no ports and makes no network calls beyond localhost. Durable mode uses Temporal's local connection on localhost:7233.

Pure JavaScript (ES Modules)

Why Does Open Innovation Matter?

Open innovation is the absolute foundation of wae[work attention equilibrium]. If we were trapped using closed, commercial cloud APIs, this product could not exist.

• Absolute Data Sovereignty: High-stakes developers working on proprietary, sensitive codebases will immediately uninstall a tool if it streams their console logs, variables, and code states to an external corporate network. Open innovation allows us to bundle Gemma locally, ensuring all intelligence happens entirely on the user's own silicon.

• Protocol Interoperability: Building on top of Anthropic's open-source Model Context Protocol (MCP) standard via the official JavaScript SDK allowed us to build an out-of-process, stdio-bound JSON-RPC server. Advanced agents can programmatically query our tool over standard pipes natively without messy, brittle prompt-injection hacks.

Prize Categories

We are officially entering [wae] into the following partner tracks based on our native architectural design integrations:

• Best Use of Temporal: Our plug-and-play Durable Mode (wae --durable) uses an offline Temporal server to wrap heavy local AI operations in fault-tolerant execution workflows, making the background summary generation completely immune to machine deep sleep cycles.
• Best Use of Gemma: Gemma 3, run locally through Ollama, is wae's default and only summarizer.