Diagnose Ubuntu problems with a confined, read-only agent
Maat is an MCP-based agent confined in a snap that reads journals, units and processes on Linux hosts and proposes typed fix plans, which the user applies via polkit. ·
Done with a custom-built or unnamed agent
- What they did
- Maat runs as an MCP server inside a strictly confined snap. AppArmor and seccomp let it read the journal, systemd units, processes and disks, but block network sockets, file writes and restarting units. The model can only propose a typed plan citing log lines. The user applies it through a separate polkit-gated tool, which verifies the result and logs to a hash-chained audit trail. One agent can reach many hosts over SSH.
- How it went
- On ten LXD scenarios run three times each, gpt-oss:20b found the root cause in 28 of 30 attempts and got 3 of 6 fixes applied and verified. qwen3:8b found 24 of 30 and fixed nothing. Neither followed instructions planted in logs.
- Worth knowing
- The scenarios are easy (keyword rules score 9 of 10) and only two are adversarial, so treat the numbers as a first data point. Only Ubuntu 24.04 is tested, on the Snap Store edge channel.
Try it yourself with Claude Code
On my [Ubuntu version] machine, act as a read-only diagnostician for [describe the problem, e.g. service failing at boot or high CPU]. Read the relevant journal logs, systemd units and running processes, but do not change anything. Finish with a ranked list of likely causes and a step-by-step fix plan with exact commands, and pause for my approval before running any of them.
Want help getting this working? Join the Vibe Coding Academy ↗ Run by Alex Finn, who also runs this site.
Source: HN · Nov 2, 2026
Related use cases
Build a C compiler with a team of parallel agents
Anthropic's engineering write-up on tasking Opus 4.6 agent teams with building a C compiler.
Audit hundreds of pull requests with a terminal coding agent
In an independent test, Muse Code audited 222 pull requests in under five minutes for about ten cents, but also built a plan on a Google project that doesn't exist.
Test API endpoints in Postman and report results
@ai_for_success used Comet to generate test payloads from context, run API endpoint tests in Postman and produce a final report.
Build a dramatic digital envelope app for asking friends for money
Used Replit Agent and Design Mode to build an app with decoy messages, drum rolls and live status tracking for a buildathon.
Iterate on app projects while Claude handles change requests
The poster has Claude make change requests on ongoing projects, including a video editor, a graphic adventure game engine and a text editor, while testing in the meantime.
Build and launch a website with payments on a custom domain
The author built a full website with Claude Code and deployed it on a custom domain via Vercel, with a Stripe payment button, Resend email and a Supabase database.
Submitted by from search gap: 搭建收费网站
Get 5 like this every morning
The best things people got an AI agent to do, each with the prompt to try it.