
Article 4/20 in the AgentKit v3 onboarding series. About a 9 minute read.
"Does AgentKit run on Cursor?" The honest answer: it installs, but it is not certified yet. This article gives you the table I use internally to decide how "mature" each runtime is, then shows how to install for each case.
1/ The v3.0.0 runtime certification table
Each runtime is trial-run on 3 environments: macOS arm64, Windows 11, Linux (WSL). Results:
| Runtime | ID | macOS | Windows 11 | Linux/WSL | Status |
|---|---|---|---|---|---|
| Claude Code | claude-code |
✅ | ✅ | ✅ | ✅ Certified 3/3 |
| DeepSeek Harness | dsh |
✅ | ✅ | ✅ | ✅ Certified 3/3 |
| Codex | codex |
✅ | ⚠️ partial | ✅ | ✅ 2/3, Windows is still partial |
| Oh My Pi | omp |
✅ | ✅ | ✅ | 🧪 Spike stage, 3/3 pass |
| Pi | pi |
✅ | ✅ | ✅ | 🧪 Spike stage, 3/3 pass |
| Cursor | cursor |
⚠️ | ⚠️ | ❌ | 🧪 Not passing yet (2 partial, Linux fails) |
| Antigravity | agy |
⚠️ | ⚠️ | ⚠️ | 🧪 Spike, 3 partial |
| Grok | grok |
- | - | - | 🧪 Kit package exists, not measured |
| OpenCode | opencode |
- | - | - | 🗺️ Not measured |
| CloudHarness | cloud-harness |
- | - | - | 🧪 Kit package exists, not measured |
| Portable (export skill ZIP) | portable |
- | - | - | 🧪 File export only, not measured |
How to read this table?
If you use Claude Code or DeepSeek Harness: relax, this is the main road.
Codex: works well on macOS and Linux. On Windows it runs, but not everything is green.
Cursor, Antigravity: installable to try out, but do not take them to important work right now.
One thing I have to admit: the agentkit.best FAQ page currently lists Antigravity as an "available" runtime, while the certification table says spike andak updatedoes not acceptagyas a target. I logged it in the gap report so it gets fixed to match.
2/ Project or global?
| Project | Global | |
|---|---|---|
| Installs into | The current project folder | The runtime's user folder (for example ~/.claude) |
| Who sees it | Only that project | Every project on the machine |
| Can be committed to git | Yes, shared with the team | No |
| When to use | Default, teams, client projects | Personal machine, when you want skills everywhere |
My advice: start with project. Your team clones the repo and has every skill, regardless of what anyone installed on their own machine.
3/ Installing for one runtime ✅
By default ak kit init installs for Claude Code, in the current project:
cd ~/projects/your-project
ak kit init engineer --yesFor Claude Code, project-native mode merges the content into ./.claude and stores lifecycle metadata in ./.agentkit.
To install globally:
ak kit init engineer --global --yes4/ Installing for several runtimes at once ✅
--target takes a comma-separated list of IDs:
ak kit init engineer --target claude-code,codex --yesAdd the marketing kit to the same project:
ak kit init marketing --target claude-code,codex --yesUpdates then follow per runtime too:
ak update --target codex --yesAccording to the ak update help in 3.0.0, the supported project targets are: claude-code, codex, cursor, dsh, grok, omp, pi. Global is the same minus pi.
5/ Installing only a few skills ✅
The Engineer kit has around a hundred skills. If you do not need them all:
ak kit init engineer --skills ak-plan,ak-cook,ak-debug,ak-ship --yesOr install everything except a few:
ak kit init engineer --exclude-skills ak-motion-video,ak-filmmaking --yesOr pick with an interactive menu:
ak kit init engineer --select-skills6/ Situations you will run into
"I edited a skill file, will the update overwrite it?"
It does not overwrite silently. ak records the hash of every file it installs. A file you edited is detected as drift, and the command stops with exit 6. To overwrite you must pass --force, and it still takes a snapshot before writing.
"I want to reset a runtime cleanly back to exactly the kit content"
ak kit init engineer --fresh --target codex--fresh shows a preview, asks for confirmation, takes a verified snapshot, and only then cleans up. In a script you need --yes, and if the plan lists risks you must also pass the exact --accept-risks token it prints. --yes alone never accepts risks on your behalf.
"Another agent is running an update"
Exit 4: another AgentKit operation is holding the kit lifecycle lock. Wait a moment and run again. Two updates can never trample each other.
"Does Codex read the project rules?"
Codex project rules are written to <project>/.codex/AGENTS.md. According to the certification table, this file is kept but not yet activated by Codex. Put important rules in the repo's root AGENTS.md. Also, the privacy guard on Codex does not yet cover MCP and local read tools (see article 7).
"Can I use the Claude Code plugin instead of copying files?"
ak kit init engineer --switch-to-pluginThis flag picks the Claude Code plugin distribution mode at the scope you choose.
Exit codes worth remembering
| Exit | Meaning |
|---|---|
0 |
Success |
2 |
Wrong flag/argument |
3 |
--fresh did not run because it needs confirmation |
4 |
Another operation is holding the lock |
5 |
Kit not found |
6 |
A file was modified (drift), or a skill alias collision |
15 |
Renewal required to download this version (article 3) |
Next article: Managing skills and SkillX →
Sources
v3.0.0 runtime certification matrix in the
agentkitrepo (tagv3.0.0)ak kit init --help,ak update --help(ak 3.0.0): target list, flags--target,--global,--skills,--exclude-skills,--fresh,--accept-risks, exit code tableFAQ on https://agentkit.best (checked 02/11/2026): the "available" and "preview" runtime lists
GitHub release
v3.0.0, per-runtime kit package asset list: https://github.com/bestagentkits/agentkit/releases/tag/v3.0.0