Skip to content
ctxloom

llm-tool-killer (ltk)

llm-tool-killer (ltk) is a sibling project in the ctxloom org. Where ctxloom shapes the context an AI coding agent sees, ltk guides the commands it runs.

It’s a small static binary the agent calls as a pre-tool hook: it parses the shell command the agent is about to run and, if one of your rules matches, turns it away and tells the model what to do instead — so the agent retries the right way rather than hitting a silent failure or an opaque block.

go test ./...   ⟶   ✗ "Run tests through the task runner."   →   the agent retries with `just test`

Why it pairs with ctxloom

  • ctxloom weaves the right context into the session — fragments, profiles, prompts, memory.
  • ltk keeps the agent on your rails once it starts acting — steering it toward your task runner and away from commands you’d rather it not run, with the suggestion you provide.

Together, the agent gets the right context and runs the right commands.

What it does (and doesn’t)

ltk parses and understands the command — resolving known variables and re-parsing trivial wrappers — then matches rules against the real command, however it’s dressed up. That’s far more robust than a substring/regex denylist that misfires on quoting or gets bypassed by a sub-shell or eval.

It is a cooperative redirect, not a sandbox. If you instruct the agent to work around a rule, it can. For hard “never, under any circumstances” isolation, run the agent in a container — ltk makes the easy, accidental path the right one.

Install

expressiveCode.terminalWindowFallbackTitle
go install github.com/ctxloom/llm-tool-killer/cmd/ltk@latest
# or build a static binary from source: just build-static

Register the hook and scaffold a starter rules file. manage install auto-detects your agent (e.g. a .claude/ directory) and merges the hook non-destructively:

expressiveCode.terminalWindowFallbackTitle
ltk manage install         # write the agent hook + .ltk/config.yaml
ltk manage install --print # dry run: show the merged config, write nothing
ltk manage uninstall       # cleanly remove the hook again

Commit .ltk/config.yaml alongside your code. Rules are YAML; the first matching deny wins and returns your message/suggest to the model.