Give your agent eyes, not guesses.

One CLI. Many debuggers. Your AI agent stops guessing at runtime state and starts observing it.

View source Watch demo ↓
Agents guess. Debuggers observe.

The bottleneck for agentic engineering is not generation. It is diagnosis.

Without dbg

The agent reads source code, builds a theory, rewrites something, and hopes. Wrong guess? It guesses again. Five cycles of guess-and-rebuild per bug. Tokens, time, and trust — burned.

With dbg

The agent sets a breakpoint, steps to the crash, inspects the variable, and reads the actual value. Root cause on the first pass. One cycle. Done.

Observe. Reason. Fix.

The agent receives a crash report. Instead of guessing, it launches a debug session, steps to the error, and reads the state.

claude — dbg session
Two commands. That's it.

Install the binary, point it at your agent. No configuration, no wrappers, no adapters.

1

Install

curl -sSf https://raw.githubusercontent.com/redknightlois/dbg/main/install.sh | sh
# or, if you already have Rust:
cargo install dbg-cli
2

Connect to your agent

dbg --init claude    # or: dbg --init codex

The agent now knows when and how to use dbg. It picks the right backend, sets breakpoints, and interprets debugger output on its own.

Many languages, one interface.

The agent learns dbg once. It works everywhere.

Debuggers
Rust
LLDB
C / C++
LLDB
Zig
LLDB
Python
PDB
Go
Delve
.NET / C#
NetCoreDbg
Java
JDB
Kotlin
JDB
JavaScript
node inspect
TypeScript
node inspect
PHP
phpdbg
Ruby
rdbg
D
LLDB
Nim
LLDB
Haskell
GHCi
OCaml
ocamldebug
Profilers
Go
pprof
Linux
perf
C / C++
callgrind
Rust
callgrind
D
callgrind
Nim
callgrind
Zig
callgrind
Memory
massif / memcheck
Python
cProfile
.NET
dotnet-trace
Node.js
V8 --cpu-prof
PHP
Xdebug
Ruby
StackProf
Haskell
GHC profiler
GPU Profilers — gdbg
CUDA
nsys + ncu
PyTorch
torch.profiler
Triton
nsys + proton
How it works.
Daemon, not subprocess.The debugger stays alive across commands. No startup cost per interaction, no lost state between steps.
PTY, not pipes.The debugger thinks it's in a real terminal. No pipe-mode quirks, no missing output, no behavior differences from docs.
Output cleaning.Backends strip DWARF indexing, thread lifecycle, symbol loading noise. The agent gets clean, token-efficient output.
One interface, many languages.A backend trait maps each debugger's commands to a common CLI. The agent learns break, step, print once.