Get Alpha Harness for your computer.

Dangerous commands are stopped before they run, risky file writes are stopped before they land, and your rules are added to every session automatically.

Same protection as Claude Code. Commands and file writes are both checked before they happen, and your rules are sized so that every tool reading them gets all of them.

Cannot: Since June 2026 Google serves the Gemini CLI only to paid API keys and Gemini Code Assist Standard or Enterprise. Individual accounts moved to the Antigravity CLI.

Same protection as Claude Code. Codex can silently stop reading a rules file that is too long; Alpha Harness checks yours fits, so no rule is dropped without you knowing.

Dangerous commands, risky file writes and deletes, and risky MCP tool calls (add-ons that act in other apps, such as sending an email) are all stopped before they happen, with the reason handed back so the model corrects itself.

Cannot: Your notes are not added to each message: Cursor gives no way to add to a message, so only the start of a session gets context. The file-write and MCP guards are not yet proven against a live Cursor session on our machines; they follow Cursor's documented hook format (read 2026-10-29).

Dangerous commands and risky file writes are stopped before they happen, in the IDE and the agy CLI. Your rules come from the same file the Gemini CLI reads, so there are no two copies to drift apart.

Cannot: Your notes are not added to a session. Not yet proven against a live Antigravity session on our machines; the guards follow Google's documented hook format (read 2026-10-30).

Same protection as Claude Code. Hermes builds its own skills as it goes, so a guard that catches a bad command before it runs matters more here, not less.

Same protection as Claude Code. The guards are added to Kimi Code's own config file; everything else in it stays exactly as you wrote it, and uninstalling gives the file back unchanged.

Cannot: Not yet proven against a live Kimi Code session on our machines. The guards follow the hook format in Kimi Code's own source (read 2026-10-23).

Dangerous commands and risky file writes are stopped before they happen. The guards live in a file of their own, so nothing of yours is edited.

Cannot: Your notes are not added to each message: Copilot CLI drops what a prompt hook says, so only the start of a session gets context. Not yet proven against a live Copilot CLI session on our machines; the guards follow GitHub's documented hook format (read 2026-10-23).

Dangerous commands and risky file writes are stopped before they run. Your rules reach it through the file it already reads, so nothing you wrote for it stops being read.

Cannot: OpenCode 2 runs no add-on written for OpenCode 1, so there it blocks nothing; the Doctor tells you if yours is 2. Your notes are not added to each message. Not yet proven against a live OpenCode session on our machines.

Your conventions are added to Aider's own config so they load in every session. Existing settings in that file are merged, never replaced.

Cannot: Aider has no hook system, so nothing can be blocked.

Your rules are added to Droid's own instructions file, so they load in every session. What you wrote in that file stays exactly as it is.

Cannot: Droid has hooks that can stop a command. Alpha Harness does not install them yet, so here it adds your rules but blocks nothing.

Your rules are added to the instructions file Amp always includes, so they apply in every project without copying anything into each one.

Cannot: Amp can stop a command through a plugin. Alpha Harness does not install one yet, so here it adds your rules but blocks nothing.

Your rules are added to Pi's own instructions file. If you moved Pi's folder, they go where Pi looks; if another file there would be read instead, the installer tells you.

Cannot: Pi can stop a command through an extension. Alpha Harness does not install one yet, so here it adds your rules but blocks nothing.