Claude Code
- config ·
- ~/.claude.json
- format ·
- JSON
- plugins ·
- yes
- dormancy ·
- entry stored and restored
Claude Code, Codex and OpenCode each store MCP configuration somewhere different, in a different format. AgentPack writes one capability into all three, launches every server to confirm it actually answers, and can undo the whole run.
Unsigned, so SmartScreen warns once: More info → Run anyway.
SHA-256 loading…
31 seconds in the built-in demo sandbox: scan, install into all three agents, verify, roll back. Recorded from the real app.
Seven stages. Nothing is reported as working until it has been observed working.
Finds installed agents and reads your project stack, with the file that proves each signal.
Deterministic rules, not a model. You can read them and predict the output.
One capability definition becomes JSON, TOML and JSONC, each written surgically.
Configs are backed up first. Credentials are requested only where declared.
Each server is launched and asked for its tool list. Not "we wrote a file".
Switch a capability off without uninstalling it. Context cost is measured, not guessed.
Restores original bytes, keeps edits you made afterwards, refuses when it cannot prove ownership.
Adding a fourth is an adapter, not a rewrite.
What it needs.
--demo for a sandbox.Stated, not hidden.