AI Extensions: one mouse mover, one config rewriter, one email box, one license flag
Four open-source extensions from the past two weeks, source-checked on September 24, 2026, with setup notes, licenses, dated activity snapshots, security caveats and one license warning.
AI EXTENSIONS · DAILY DIGEST
Three of today's projects would make a careful person pause. One moves your real mouse and keyboard. One rewrites the config files of every AI tool it finds on your machine. One gives agents a working email address. All three are useful, and all three deserve a read before you install. This digest covers four open-source extensions from the past two weeks, plus one license warning.
QUICK ANSWER
Four new projects worth inspecting, none worth installing blindly
A cheap computer-use loop built on a decision model, a plain-text file format for assistant memory, a source-tracing study skill and a self-hosted email inbox for agents. Each links to its official repository with setup notes, license, a dated activity snapshot and a security review.
How this digest was put together
- I read each project's official README and license on GitHub and took star, fork and last-push figures from the GitHub API on September 24, 2026. Stars are cumulative and easy to inflate; last push shows recency, not good governance.
- Not ranked. Inclusion means “worth inspecting.” I did not install or run any of these projects.
Computer use / decision model
typesafe-computer-use
What it does and why it matters
Give it a goal in plain English and it steers a Mac toward it. It reads the screen with OCR, asks TypeSafe's Jev decision model to choose the next action from a short list, and only calls a writing model when a field needs free text or the classifier gives up. The author reports about $0.0002 per decision against $0.032 for a frontier model on the same screenshot. The README is straight about the price of that saving: reasoning a big model does for free, like reading dates, has to be rebuilt as deterministic code. It is a working example of the hosted decision-model pattern we covered yesterday.
Setup / use
macOS 14+, Python 3.12+ and uv. Clone, run uv sync, copy .env.example to .env and add a TYPESAFE_API_KEY (an Anthropic key is optional). The terminal needs Screen Recording and Accessibility permissions. Windows support is experimental. The project is labeled beta.
Popularity and maintenance snapshot
939 GitHub stars · 76 forks · created 2026-09-16 · last push 2026-09-23. Observed 2026-09-24; popularity is not a quality score.
License
MIT.
Security and privacy review
This tool clicks and types on your real desktop and sends screen text to a hosted classifier at every step. Start with a dry run, use a separate macOS user with no saved passwords, and pull the Accessibility permission when you finish. The README's example task ends at a checkout page. Never let it complete a payment unattended.
Memory / file format + MCP server
.dai (daidocs)
What it does and why it matters
An open format that stores assistant memory as plain-text files on your disk: a YAML header, a JSON block and the text. A cheap “observer” model writes the store once and any model can read it, through an MCP server with save_memory and recall_memory tools. The bet is that memory should be a file format, not a vendor database. The project claims second place on the LongMemEval-S leaderboard among re-runnable systems and ships per-question judge verdicts with a sha256 manifest. That is the project's own benchmark claim, which I have not reproduced.
Setup / use
Node 18+. npx daidocs setup detects Claude Desktop, Claude Code, Cursor, Windsurf, Codex, Cline, Continue and Zed, configures them, installs session hooks and backs up the files it changes. You can also clone the repo and run node setup.js.
Popularity and maintenance snapshot
40 GitHub stars · 3 forks · created 2026-09-13 · last push 2026-09-24 · 16 open issues. Observed 2026-09-24; popularity is not a quality score.
License
Apache-2.0.
Security and privacy review
Plain text is easy for you to inspect, and just as easy for anything else on the machine to read. Keep the store out of synced or shared folders unless you mean it, and keep secrets out of it entirely. The one-command setup edits many tool configs and installs hooks, so read the diff of the backup it makes, or try setup in a test account first. Your conversation history goes to whichever observer model you configure.
Agent skill / learning
learn-from-materials
What it does and why it matters
Feed it books, PDFs, slides and web pages and it builds a traceable knowledge base, an offline study page and synced Markdown. What makes it worth a look is source tracing: claims map back to PDF pages, slides or EPUB chapters, and the output keeps material-backed facts, gaps in the material and model-added content apart. It also runs coverage audits and keeps your notes in the local browser. It is an Agent Skills package.
Setup / use
Python 3.10+. Clone the repository into your agent's skills directory, for example ~/.claude/skills/learn-from-materials for Claude Code or ~/.agents/skills/ for Codex, then ask the agent to use it on a file. Node, Playwright and Chromium are optional for browser checks.
Popularity and maintenance snapshot
462 GitHub stars · 40 forks · created 2026-09-08 · last push 2026-09-21. Observed 2026-09-24; popularity is not a quality score.
License
MIT for the skill. The README notes that sample source texts and screenshots remain subject to their rights holders.
Security and privacy review
The skill runs local Python on the documents you hand it. The README says it works offline by default and does not install dependencies on its own. Check that before you rely on it, and remember your agent's model still reads the material. Keep confidential documents out unless your model provider's terms allow it.
Agent infrastructure / email + MCP
DearAgent
What it does and why it matters
Run your own inbox API for agents on your own Cloudflare account. Agents can open catch-all inboxes, wait for a verification code, reply in thread, send mail and receive HMAC-signed webhooks. Mail sits in D1, attachments in R2, and an MCP endpoint at /mcp exposes the same tools. It pitches itself as an open alternative to hosted inbox services.
Setup / use
A Cloudflare account, an apex domain that receives no other mail, Node 20+ and npx wrangler login. Clone, run npm install, then npm run setup for the interactive deploy. Receiving works on the free plan; sending needs Workers Paid.
Popularity and maintenance snapshot
27 GitHub stars · 3 forks · created 2026-09-11 · last push 2026-09-14. Observed 2026-09-24; popularity is not a quality score.
License
AGPL-3.0. If you modify it and offer it as a network service, the AGPL's source-sharing terms apply.
Security and privacy review
The author calls it experimental and says it has had no in-depth audit. It is single-tenant with one API key, so any agent holding that key can read every inbox. Verification codes and password resets are high-value mail. Give it a dedicated domain, keep retention short and do not let agents use it for accounts that matter.
The license flag
anything2explainer, a Claude Code and Codex skill that turns a topic into a narrated explainer video, was the most-starred new agent-skills repository in today's search (2,025 stars, observed 2026-09-24). GitHub shows its license as unrecognized. The README says it uses PolyForm Noncommercial: free for noncommercial use, with commercial use needing the author's prior permission. That is not an open-source license, and it matters if you plan to use it at work.
Analysis (Jack Gross)
My reading, not a reported fact: three of today's four projects widen what an agent can touch. That is your desktop, every tool config on the machine, or an email identity. The READMEs worth copying are the ones that state their own limits: a beta label, a missing audit, a benchmark caveat. A project that skips those notes deserves more caution, not less.
FAQ
Are these ranked?
No. They are new projects worth a look, chosen for variety and clear documentation.
Did you test them?
No. Nothing was installed or run. Every claim about cost and benchmarks comes from the project's own README.
Which is safest to try first?
The study skill, because it only runs local Python on files you pick. The computer-use and email tools call for an isolated account.
typesafe-computer-use · daidocs · learn-from-materials · DearAgent · anything2explainer