assess wraps your commands, catches errors, and auto-suggests fixes using AI. Built for Linux and macOS — Windows coming soon. No telemetry, no background daemons, no compromise.
Not a chatbot. Not a wrapper around ChatGPT. A focused CLI that intercepts failures at the moment they happen and gives you a one-key decision: run the fix, or don't.
When a command fails, assess sends the error + context to Groq and gets back a corrected command — typoes, missing flags, wrong syntax all auto-fixed.
Safe → standard prompt. Soft Warn → sensitive ops (curl|sh, sudo) show the URL and require typing "yes". Hard Block → truly destructive ops refused outright.
assess ask "list files larger than 10MB modified today" returns the right find command, with OS / shell / CWD / available tools sent as context.
Every accepted fix is logged to ~/.assess_history with timestamp. Every executed command (original + fix) is appended to ~/.assess_audit for incident review.
If an AI suggestion also fails, the new error is fed back automatically for a second attempt — depth-limited to 1 to avoid loops.
Detects &&, ||, |, >, $() and routes through /bin/sh for proper parsing. Simple commands still use wordexp for correct path quoting.
API key stored with mode 0600. Dry-run mode refuses to execute anything. Sudo commands get explicit confirmation. No telemetry.
Pure POSIX C with vendored cJSON — no -dev packages needed. Runs on Linux and macOS today. Native Windows build coming in v2.0.
Most "AI safety" wrappers either block everything useful or block nothing dangerous. assess draws the line in the right place — install scripts work with explicit consent, rm -rf / never runs.
Normal commands — ls, cp, git, etc.
Sensitive — install scripts, sudo. Shows URL, requires "yes"
Destructive — rm -rf /, dd to device, mkfs, fork bombs
Unquoted globs like rm -rf * are expanded by your shell before assess sees them. Quote globs to let assess catch them.
Instead of running mkdirr dir, run assess mkdirr dir. assess forks, executes, and captures stderr.
Common errors (file not found, permission denied, command not found) are handled locally without an API call — instant response.
For typoes and unknown errors, assess calls Groq with the command + stderr + system context. You see analysis + suggested fix, then choose to execute or skip.
assess ask sends your OS, shell, current directory, and detected tools as context to the AI — so suggestions are accurate for your machine, not a generic Linux box.
No package manager required. The vendored cJSON source means you only needlibcurland a C compiler. Works on Linux and macOS today — Windows is coming soon.
git clone https://github.com/aadicodesforfun/assess.gitcd assessmakesudo make installassess -s gsk_your_key_hereassess --set-model llama-3.1-8b-instantassess ask "list files larger than 10MB"macOS ships with everything except libcurl headers. Install via Homebrew:
brew install curlThe button auto-detects your platform. Linux gets the 77 KB executable; macOS gets a zip with all source files to build locally.
On Windows? Native support is coming in v2.0 — for now, use WSL2 to run the Linux build.