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.

Notes

• 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.

Features

• 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)

How to use

Install

Configure

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.

Required fields

• LLM provider (e.g. Anthropic)
• Model (e.g. claude-haiku-4-5)
• API key (e.g. JQ.Ab9RN6W6QW7cmcnY92DIuoVtjCpKm_qfmO5T5oGzQmnwe5fjhw)

Notes

• 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.

Additional required field for custom model only (Optional override for Ollama/Enterprise endpoints)

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

Use

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

Chaining Multiple Commands

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"

Use in agentic workflow

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`).

Outcome

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

Example output

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

                                                     🧠📜 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

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

CLI Flags

Environment Variables

How It Works

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 (exit 0, no warning-like keywords): reports success and sends nothing to the LLM.
• Success with warnings (exit 0, keywords such as warning, deprecated, or skipped 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)"]

Notes

• 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 run brytlog --config to open it.
• Logs are saved locally to brytlog-reports/ and brytlog-raw/ in the current working directory. You may want to add brytlog-*/ 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.

License

MIT