MoonshotAI
Featured
Kimi-K3
Kimi K3 is an open-weight, 2.8T-parameter native multimodal agentic model with a 1M-token context window, designed for frontier coding, knowledge work, and reasoning tasks.
AI repository intelligence
FR-AI / ANALYSISWhy it is worth attention
It is a practical, real-world automation that combines screen recording, AI agent orchestration, and personal productivity into a daily habit, with a privacy-first design and a feedback loop that refines future audits.
Who it is for
- Developers building AI agent workflows and automation
- Productivity enthusiasts who track their digital activity
- macOS users familiar with launchd and CLI tools
- Users of Coast screen tracking looking for advanced recap capabilities
Use cases
- Daily morning review of unfinished communications and work-in-progress
- Tracking follow-ups on messages, emails, or tasks left incomplete
- Automatically documenting and archiving daily loose ends with visual evidence
- Building a personal habit of accountability and closure
Strengths
- Uses vision-based verification to ensure screenshot crops actually show claimed evidence
- Runs two subagents in parallel to cover both communication and work-in-progress loose ends
- Privacy-sensitive: all sensitive data remains on local machines and is gitignored
- Integrates push notifications, a web UI, and a two-way sync for seamless interaction
Considerations
- Tightly coupled to Coast.app, OpenCode, and macOS (launchd), making cross-platform adoption difficult
- Requires non-trivial setup including multiple CLI tools and a private hosting box
- Resource overhead from screenshot cropping, OCR, and vision analysis may be significant on weaker machines
README quick start
coastcap — morning audit
Every morning, an agent reconstructs yesterday from my Coast screen history, finds the loose ends — things I left unfinished, never sent, or forgot to follow up on — and publishes a recap page with cropped screenshot evidence. A push notification lands on my phone when it's ready. Marking an item complete on the page feeds back into the next morning's audit.
(full quality: docs/demo.mp4)
How a morning runs
- 9:00 — launchd starts a headless OpenCode run
(
run.sh+prompt-template.md) with a random jitter. - The agent reads its memory: previous audits (
history/), the previous recap JSON, and which items I marked complete on the page (state/interactions.json). - It dispatches two subagents in parallel over the Coast CLI: one hunts communication loose ends (unsent replies, people waiting on me), the other work-in-progress loose ends (abandoned terminals, half-typed prompts, errored flows). Carryover items from previous days get re-verified against the new day's evidence.
- For the items where a screenshot helps, it locates the exact frame, crops
to the evidence region using Coast's OCR box coordinates
(
scripts/crop-evidence.py), then looks at each crop with vision and only ships images that actually show the claimed evidence. - It writes
site/data/.json. The Bun server renders it as a per-day page (/july-28th-2026) with an archive at/. - A 10-minute two-way sync (
scripts/sync.sh) pushes the site to an exe.dev box (auth-gated HTTPS) and merges item interactions in both directions. The first sync after a new recap fires a Hark push notification with a tap-through to the day's page. - Tapping a row on the page blurs it and offers Mark complete — that interaction syncs back and becomes "cleared" in tomorrow's audit.
Components
| Path | What it is |
|---|---|
run.sh | Morning entrypoint: date substitution, jitter, headless opencode2 run --auto |
prompt-template.md | The agent's brief: memory, subagents, synthesis rules, crop verification, output schema |
scripts/crop-evidence.py | Frame + on-screen text → OCR-box union → padded sips crop |
scripts/sync.sh | Two-way state merge + site push to the box + once-per |
Related repositories
Similar projects matched by category, topics, and programming language.
jamesob
Featured
local-llm
A comprehensive guide for building and configuring a high-end local machine to run state-of-the-art LLMs, with detailed hardware choices, BIOS tuning, and Docker-based model serving.
Blaizzy
Featured
nativ
Nativ is a native macOS app that lets you run AI models locally on Apple silicon, offering chat, model management, performance analytics, and an OpenAI/Anthropic-compatible API server.