Show HN: Brytlog – AI logger

Brytlog replaces raw logs with AI summary, thus saving developers time, trouble and money.

In agentic workflows brytlog acts as a cheap, fast pre-processor to the chief agent.

For example, Claude Opus 4.8 (chief agent) might run brytlog python run.py

, rather than the plain python run.py

. This way, instead of having to process the entire raw output on its own (slow, expensive, bloats context), it will only get a concise summary, generated by a cheaper, faster model, such as Gemini-3-flash.

• As a fail-safe, raw logs are saved so they can be accessed by the chief agent or developer if still needed (this is a toggleable feature in config).
• Even the cheaper model doesn't get the full raw dump, just the important parts, thus saving even more time and money.

In non-agentic, dev-driven workflows, brytlog simply saves the developer the time and trouble of analyzing raw output by himself, or copy-pasting lines into a coding assistant.

• free
• open source
• platform, language and llm vendor agnostic
• minimal setup (just bring your own key, or run locally)
• no need to change existing code (just add a couple of lines to AGENTS.md)
• lightweight (~50 KB, ~1,400 lines of code)
• customizable
• privacy-minded (brytlog doesn't collect any data, and it redacts sensitive information before passing it to the LLM)

demo-no-brytlog.mp4 | #

demo-with-brytlog.mp4 | #

demo-json-output.mp4 | #

pip install brytlog

After installation either run a command using brytlog (e.g. brytlog node main.js

), which will launch an on-boarding config process in the terminal, or run brytlog --config

to open a json configuration file in your default editor.

• LLM provider (e.g. Anthropic)

• Model (e.g. claude-haiku-4-5)

• API key (e.g. JQ.Ab9RN6W6QW7cmcnY92DIuoVtjCpKm_qfmO5T5oGzQmnwe5fjhw)

• Google Gemini: Use keys from Google AI Studio ( [https://aistudio.google.com/]). Google Cloud Vertex AI is not natively supported yet.- Custom Providers: Supports OpenAI-compatible endpoints (Ollama, vLLM, etc.) via standard Authorization: Bearer

headers. Azure OpenAI is not natively supported yet.

• Base URL (e.g. http://localhost:11434/v1)

Simply prefix any command with brytlog

.

Syntax: brytlog [options] <command>

.

Examples:

brytlog python run.py 
brytlog --api-key xyz... node build.js
brytlog --model claude-haiku-4-5 ./deploy.sh

Wrap the entire string in quotes to run multiple commands together.

brytlog "pip install -r requirements.txt && python run.py"

brytlog "node main.js
         npm run serve"

Either prompt inline at the beginning of a session or add this or similar to AGENTS.md:

brytlog replaces raw terminal output with a concise AI summary.

Use it for interpreters (e.g.`python`, `node`), compilers, build tools (e.g. `npm`, `make`), and test runners (e.g. `pytest`). Do not use it for standard OS utilities (e.g.` ls`, `cat`), version control (e.g.` git`), or interactive CLI tools (e.g.` htop`).

Syntax: `brytlog [options] <command>` (e.g., `brytlog --json python run.py`).

Instead of raw log, a short report is outputted to the terminal.

                                   ────────────────────────────────────────────────────────────

                                                     🧠📜 brytlog crash report

Problem
The program crashed due to a TypeError when attempting to divide an integer by a string. The 'items' value from the configuration is a string and needs to be converted to an integer for arithmetic operations.

Fix
Convert payload['items'] to an integer before division in /var/folders/ys/zfg72rwd661dn0cnrsqth5sh0000gn/T/tmph1e0wuse.py, line 6:
average = payload['total'] / int(payload['items'])

Full report → /path/to/project/brytlog-reports/2026-06-15T14-11-51.log
Full raw log → /path/to/project/brytlog-raw/2026-06-15T14-11-51.txt

                                   ────────────────────────────────────────────────────────────

brytlog runs a given command as a child process, diverts its stdout

and stderr

away from the terminal by default, and instead only outputs a concise AI summary of the run.

Possible scenarios:

Clean success(exit0

, no warning-like keywords): reports success and sends nothing to the LLM.Success with warnings(exit0

, keywords such aswarning

,deprecated

, orskipped

detected): sends a sampled head / warnings / tail excerpt to the chosen LLM for a short summary of the run.Crash(non-zero exit): sends a token-bounded head-and-tail excerpt of the output to the LLM and prints a concise problem/fix report.

Because brytlog wraps the process rather than hooking into it, it works for any language or runtime with no changes to the program being run.

flowchart TD
    A[Run command] --> B[Capture output and exit code]
    B --> C{Success}

    C -- Yes --> D{Warnings detected}
    C -- No --> F[Build excerpt]

    D -- No --> E[Print success]
    D -- Yes --> F

    F --> G[Send to LLM]
    G --> H[Print AI report]

    B -.-> I["Save raw log (optional)"]
    H -.-> J["Save AI report (optional)"]

• CLI flags take precedence over environment variables, which in turn take precedence over the config file.
• Configuration is saved globally to ~/.brytlog.json

(C:\Users\Name\.brytlog.json

on Windows). You can edit this file directly or runbrytlog --config

to open it.- Logs are saved locally to brytlog-reports/

andbrytlog-raw/

in the current working directory. You may want to addbrytlog-*/

to your.gitignore

.- To be clear: raw command output is muted (run in quiet mode) to humans and AI agents by default to save tokens and prevent context bloat. In quiet mode, stdin

is redirected to/dev/null

, meaning any interactive prompt will instantly crash the program (e.g.EOFError

) to generate a helpful AI report rather than hanging indefinitely. Brytlog only prints the concise AI report if a command fails. To display the live stream of raw logs in the terminal as usual and answer interactive prompts, use--no-quiet

.- Reasoning models (e.g., Gemini Flash-3-Preview): These models spend a large portion of their output budget on "thinking" tokens. Brytlog automatically enforces a 2048-token floor at the API level to ensure these models have room to think, while still instructing them to keep the visible report under your MAX_OUTPUT

limit.- Enforcing agent use of brytlog, rather than relying on AGENTS.md or direct prompting, is also possible (e.g. via subprocess shim on PATH), but out of scope in this version.

MIT