AI Extensions: three tools with three kinds of power, and no reason to trust a star count
Three new open-source projects checked September 25 for setup, license, repository activity and security exposure. Stars are observations, not endorsements.
AI EXTENSIONS · DAILY DIGEST · 25 SEP 2026
Today’s three extensions hand out very different powers. One reads personal conversations. One judges evidence through a hosted model. One gives a chat agent local command execution. A promising star count is not a reason to install any of them.
What it does and why it matters
- I checked the official README, LICENSE and GitHub repository API for each project on September 25, 2026. Stars and forks are cumulative snapshots, not quality scores or growth rates. I did not install or run any of these packages; functions and setup below are project documentation claims, not independently tested results.
Relate
What it does and why it matters
A local hub sorts conversations into separate spaces, pulls out promises still pending, offers reminders and gives each space its own MCP endpoint. Its value is a boundary you can inspect between chat ingestion, memory and a draft a human approves; it is not evidence that inferences about feelings are correct. README.
Setup path (as documented)
Windows: download RelateSetup.exe from the project Releases and configure an AI provider; macOS/Linux: Node.js 22+, clone, npm install, npm start, then use localhost:5080/hub. These are documentation steps, not commands we ran. README.
Maintenance and popularity snapshot
GitHub API, 2026-09-25: 103 stars, 0 forks; latest push 2026-09-23 16:34 UTC. Repository. Cumulative signals, not quality scores.
License
AGPL-3.0 · LICENSE (repository-level; dependencies and connectors may differ).
Security and privacy
The app reads sensitive chats, promises and mood inferences. The repository says storage stays local except for text sent to the model provider you choose. Before you connect any real conversation, review connector permissions, the HTTP bridge, plugin scope, provider retention and backups. Its MCP view is described as read-only, but a README privacy claim is not an audit.
jev-judge-mcp
What it does and why it matters
An MCP server exposes eleven typed judgment tools for verification, triage and routing. It keeps probabilistic model answers apart from explicit policy, which helps where the options can be listed and the output is not open-ended text. README.
Setup path (as documented)
Requires Python 3.12+, uv, a POSIX system and a TypeSafe API key. The documented path uses uvx with the jev-judge-mcp[typesafe] package: setup verifies and stores the key; install offers to edit agent configurations; doctor checks offline. Review its install --dry-run plan first. We did not run it. README.
Maintenance and popularity snapshot
GitHub API, 2026-09-25: 29 stars, 3 forks; latest push 2026-09-25 04:19 UTC. Repository. Cumulative signals, not quality scores.
License
MIT · LICENSE (repository-level; dependencies and connectors may differ).
Security and privacy
Judgment prompts and evidence go to the TypeSafe API. A probability plus an auto/review/escalate policy is not proof of truth and not a security boundary. The installer can change several agent configuration files, so scope it to one client and keep keys out of shared desktop config where you can.
openchatx-mcp
What it does and why it matters
It puts local shell, files, optional tools, MCP servers and subagents behind one ChatGPT custom MCP connection. The trade is plain: richer agent reach for a much larger local blast radius. README.
Setup path (as documented)
The README lists macOS or Windows 10/11, Node.js 22.18+, npm, ripgrep, OpenAI tunnel-client and ChatGPT Developer Mode/custom MCP access. Its manual path clones the repo, runs npm ci and npm run setup -- --config-only, then configures a Secure MCP Tunnel before npm run setup and npm start. None was installed here. README.
Maintenance and popularity snapshot
GitHub API, 2026-09-25: 27 stars, 1 forks; latest push 2026-09-24 19:02 UTC. Repository. Cumulative signals, not quality scores.
License
MIT · LICENSE (repository-level; dependencies and connectors may differ).
Security and privacy
The repository warns that the server inherits the OS user’s permissions, so authorized callers can run commands, edit files and drive connected tools. The tunnel and runtime keys add a sensitive authentication surface. Do not switch on broad tool access on a daily-use machine without isolation, limited credentials and a clear check of what data reaches the models.
Analysis (Jack Gross)
The three projects stretch an agent's reach in different directions: intimate chat history, a remote judgment API and local OS execution. My view is that you should scope the capability before you test the convenience. Start with synthetic data and one client, read the exact config diff, and deny network or filesystem access the experiment does not need.
How to evaluate safely
A practical test: prepare a fake conversation, a fake patch and a disposable local directory. Ask what each tool can read and write, check its outputs and logs for data leaving the boundary, then revoke access. This is a workflow I propose, not a test I ran.
FAQ
Are these ranked by stars?
No. Stars are dated snapshots. They say nothing about quality or safety.
Were the packages run?
No. The setup notes come from each project's official documentation.
Can I use them with private data?
Only after you have checked scopes, keys, retention and where each tool sends or writes data.
Sources: Relate · jev-judge-mcp · openchatx-mcp. GitHub API snapshot 2026-09-25.