Read-only tools that show what AI coding agents do to your machine — and one screen that gathers them.
Small desktop tools for developers and vibe coders who run AI on Linux. Each one answers one question about your machine, shows its evidence, and never acts in your place — you read, you decide, you act with the host's own settings or a command you have read. Want one of them? Install it, it is complete on its own. Want several? Install the control tower, then pick the modules you need.
| Verb | What it answers | |
|---|---|---|
| Permitt | inspect | What each AI assistant may do without asking — run commands, edit files, reach the network, call MCP tools — and which setting, in which file, says so. 16 hosts. Shell start-up files checked for bypass flags. |
| MCProster | inventory | Every MCP server declared on the machine, once, with the hosts and files that declare it — duplicates, secrets in clear, missing commands, skipped confirmations, what each server costs in the model's context. |
| Cleanator | clean | What AI tooling left on the disk — model stores, caches, build outputs, abandoned environments — and what could go, as a command you read and run yourself. Never deletes anything. |
| Linstruct | lint | Your AI instruction files — CLAUDE.md, AGENTS.md, .cursor/rules, copilot instructions, GEMINI.md… — as each agent will actually read them: what it cuts past its budget, what it never reads, what another file shadows, what costs context twice. 19 hosts, every rule sourced. |
| Billfold | tally | The accounts the machine is wired to and how each pays, the provider keys it holds and who can read them, the tokens each host burned in the last 30 days — priced at API rates from an embedded snapshot, never fetched. 21 hosts, 22 providers. |
| Lastrun | replay | What your AI agents actually did on the machine, session by session — commands run, files created, edited and deleted, packages installed, refusals — read back from the transcripts the hosts already write, never their text. 21 hosts, every rule sourced. |
Blackberrylabs — one screen that says what is wrong on your machine on the AI side, most serious first — act now · review · good to know · information — with every module's findings in one list. It works with any subset of the modules and shows the others as available.
Get it — .deb, .rpm, AppImage.
- Read-only. No file is written outside the tool's own preferences (and, for the tower, its saved reports).
- No network. Nothing leaves the machine; links are shown, never followed.
- Secrets masked the moment they are read.
- Actions shown, never run. Each finding says what you could do; you do it.
- Every rule sourced. The documentation sentence, the source line, or the observation behind each claim is in the tool.
Three shared crates, all Rust, all MIT: bbl-hostrules (where each AI assistant keeps its files, how to read them without writing), bbl-ui (the shared face), bbl-report (the contract between the tower and its modules). Adding a module to the family = a binary that answers report and one declarative sheet in the tower; start from the template.
By Alexandre Do. Issues and rules welcome in each repository.