a local system, watched
Your whole local AI context — under your control
onplate scans Claude, Codex, GLM and OpenCode, ties scattered skills, agents and MCP servers into one map — and shows what takes up context, what goes unused, what is out of date and what needs attention.
An example onplate report
A minute is enough to see what needs attention
skills unused for 30 days
security risk
outdated install
The numbers above are an example. onplate builds this report from your own local configuration.
system workflow / 01—03
From scan to order in three steps
The first payoff comes without moving your existing setup by hand.
- discover
Scan the system
onplate checks your local AI configs and finds the skills, subagents and MCP servers already installed. Nothing has to be moved by hand.
- understand
See the whole picture
The dashboard shows context weight, usage over 30 days, risks, outdated versions and changes made outside onplate.
- control
Fix it safely
Remove the excess, update installs, accept outside changes or roll back to a previous version. Orbi confirms the outcome or tells you what to re-check.
control surface / live
What onplate keeps under control
Not a list of technical features, but answers to the questions that matter about your local AI context.
How much context is taken
Estimates the tokens that global skills and subagents add to Claude, Codex, GLM and OpenCode, and shows the heaviest of them.
Whether a skill works at all
Hands the skill to your Claude Code, Codex or OpenCode on an ordinary task in a temporary copy, then reads the session log to show whether the model actually reached for it. Not the model’s opinion of itself — a fact you can check.
What is actually used
Reads the AI logs available on your machine and shows which skills fired in the last 30 days and which are installed but have never triggered.
What carries risk
Looks for prompt injection, signs of secret leakage, dangerous commands, suspicious URLs and hidden Unicode before you install or update.
What changed or went stale
Compares the managed state with the files on disk and finds outside edits, installs that disappeared, conflicts and available updates.
How it is all laid out
Manages skills, subagents, MCP servers and packages from one place — globally or only for the project you chose.
What exactly changed
Orbi confirms the result against the local state and warns you when it stayed unconfirmed. Versions, trash and the activity log get you back.
Honest compatibility
What each AI supports
Skills work in all four tools. For the other kinds onplate offers only the targets that really support them.
| Capability | Claude | Codex | GLM | OpenCode |
|---|---|---|---|---|
| Skills | Supported | Supported | Supported | Supported |
| MCP servers | Supported | Supported | Supported | Not supported |
| Subagents | Supported | Not supported | Supported | Supported |
| Commands | Supported | Supported | Supported | Supported |
| Hooks | Supported | Not supported | Supported | Not supported |
| Rules / instructions | Supported | Supported | Supported | Supported |
| Skill run | Supported | Supported | Not supported | Supported |
Global and project installs depend on what the particular AI can do. onplate does not offer incompatible options. A run needs the tool’s command line: GLM has none, so that row is honestly empty.
Local-first, no small print
What stays local, and when the network is needed
onplate needs no account and no cloud backend. External sources do go over the network — below is an honest split of what happens on your machine.
Analysis — on your machine
Scanning configs, estimating context and parsing local logs all happen on your computer. Session contents are never uploaded to onplate.
Network — for external sources
The connection is used for Git, marketplaces, update checks, downloading files and viewing remote resources.
Secrets are not published
Tokens and secret fields are stored locally in encrypted form and never end up in exported packages or team profiles.
The security scanner works by deterministic rules: it helps you find the suspicious, but it does not replace a manual review and is not a guarantee of safety.
First launch and the main actions
Start with sync: onplate finds what is already installed in your AI tools. Then decide what to manage, what to check and what to remove.
Step 0. Sync what you already have
Open “Sync” and press “Scan system”. onplate checks the AI home directories (~/.claude, ~/.codex and so on) and your projects, finds the skills already installed and shows their status: what can be imported, where a conflict is, and what is already managed. Press “Import all” — from there onplate keeps an eye on them.
Check that a skill actually works
The “Try for real” button hands the skill to your Claude Code, Codex or OpenCode on an ordinary task — in a temporary copy, installing nothing. The verdict comes from the session log, not from what the model says about itself: you see whether it reached for the skill, at which step, and what it cost. Pick a cheaper model and set a spending cap — what is being tested is the skill description, not the strength of the model.
And what changed since the last check
The “Runs” section keeps the history: which skill, on which version, on which task and how it ended. Repeat the same task after an update and the regression is right there: “fired on 1.3.0, stopped on 1.4.0”. The same result on the same version is not called a regression — models are not deterministic, and we will not pretend otherwise.
- 1
Import your first package
Open “Marketplace” → “Add” and paste a link to a Git repository with skills. onplate downloads it, shows the progress and sorts the contents by kind (skills, commands, MCP).
- 2
Install into your tools
In the install dialog tick the AI tools you want and the scope (globally or for a project). After writing, Orbi checks the local bindings and shows a confirmed or unconfirmed result.
- 3
Security check
The “Security” section scans installed artifacts: injections, secrets, dangerous commands. Acknowledge findings or mark domains as trusted.
- 4
Storage and cleanup
“Skill storage” has tags, favorites, content preview, a jump into the editor and the trash. Send what you do not need to the trash instead of leaving it to clutter the AI context.
- 5
Try a skill for real
The “Try for real” button on a skill card hands it to your AI on an ordinary task. You see whether it fired, what it cost and what changed since last time — for example, whether the skill stopped firing after an update.
- 6
Versions and rollbacks
Every artifact has a version history with a diff and a rollback. Edited a skill and made it worse? Go back to the previous version in one click.
Need the details?
Every part of the app has its own documentation page: step by step, with all the buttons and hints.
Install
Pick the build for your system. Detailed commands and troubleshooting live in a separate guide.
By downloading onplate you accept the terms of the license agreement.
macOS
Unpack the archive, move onplate.app to Applications and allow the first launch in Privacy & Security in System Settings.
Linux
Unpack the archive, install WebKitGTK 4.1 if the library is missing, make the onplate file executable and run it.
Windows
Run the installer, or take the portable archive if installers are blocked by policy. SmartScreen may warn about an unknown publisher: “More info” → “Run anyway”.
Frequently asked questions
Does my data go to the cloud?
Scanning configs, estimating context and parsing logs all happen locally. The network is used for the Git sources you choose, marketplaces, updates and remote resources. Session contents and secrets are never uploaded to onplate.
Which AI tools does this work with?
Claude, Codex, GLM and OpenCode. One portable skill installs into all the tools you select at once; commands and MCP servers go wherever there is a suitable place for them.
What does “could not confirm” mean?
It does not mean the operation was cancelled. It may have changed some files or config, but onplate could not confirm the outcome against the updated local state. Refresh the source section and “Bindings”, check the paths and versions, and only then repeat the action.
How is this better than the native claude plugin and codex plugin?
onplate brings together a local inventory across several AI tools, context estimates, usage observation, risk scanning and version history. Portable skills can be installed into several tools at once.
Do I need to install anything else?
No. The app is self-contained — download it, run it, use it. No extra services or environments.
Which platforms are supported?
There are builds for macOS (Apple Silicon), Linux (x86_64) and Windows (installer and portable archive).
What does it cost?
onplate is completely free — every feature, with no limits and no payment.
ready for local scan
Take your AI context under control
A local audit, a cleanup of what goes unused and a history of changes — with no account and no cloud backend.
macOS · Linux · Windows