bjarne

AI-assisted C/C++ code generation with mandatory validation.

"Within C++, there is a much smaller and cleaner language struggling to get out." — Bjarne Stroustrup

bjarne helps that cleaner code emerge by catching mistakes before they become permanent.

What It Does

1. You describe what you want in plain English
2. AI generates C/C++ code
3. Code passes through validation gates (static analysis, sanitizers, compilation)
4. Failed gates → AI fixes automatically → re-validate
5. Only validated code is presented to you

Key Features

• Multi-provider support - AWS Bedrock, Anthropic, OpenAI, or Google Gemini
• Auto-escalation - Starts with fast/cheap models, escalates to powerful models if fixes fail
• Workspace indexing - /init indexes your codebase for context-aware generation
• Configurable validation - Enable/disable specific gates via /config
• Multi-file projects - Generates and validates header + implementation pairs
• Conversation memory - Iteratively refine code in a session

Validation Pipeline:

• clang-tidy (static analysis)
• cppcheck (deep static analysis)
• IWYU (Include What You Use - header hygiene)
• Compilation with hardening flags (-fstack-protector-all, FORTIFY_SOURCE, PIE)
• ASAN (AddressSanitizer - memory errors, buffer overflows, use-after-free)
• UBSAN (UndefinedBehaviorSanitizer - signed overflow, null pointer dereference)
• MSAN (MemorySanitizer - uninitialized memory reads)
• TSAN (ThreadSanitizer - data races, deadlocks) when threads detected

Installation

Linux / macOS

curl -fsSL https://raw.githubusercontent.com/3rg0n/bjarne/master/install.sh | bash

Windows (PowerShell)

irm https://raw.githubusercontent.com/3rg0n/bjarne/master/install.ps1 | iex

Manual Download

Download from GitHub Releases:

• Linux: bjarne_X.Y.Z_linux_amd64.tar.gz / bjarne_X.Y.Z_linux_arm64.tar.gz
• macOS: bjarne_X.Y.Z_darwin_amd64.tar.gz / bjarne_X.Y.Z_darwin_arm64.tar.gz
• Windows: bjarne_X.Y.Z_windows_amd64.zip

Extract and place bjarne (or bjarne.exe) in your PATH.

Requirements

• Container Runtime: Podman (recommended) or Docker
  • Linux: sudo apt install podman or sudo dnf install podman
  • macOS: brew install podman && podman machine init && podman machine start
  • Windows: winget install RedHat.Podman
• LLM API Access (one of):
  • AWS Bedrock (default) - requires aws configure or environment credentials
  • Anthropic API - set BJARNE_PROVIDER=anthropic and BJARNE_API_KEY
  • OpenAI API - set BJARNE_PROVIDER=openai and BJARNE_API_KEY
  • Google Gemini - set BJARNE_PROVIDER=gemini and BJARNE_API_KEY

Quick Start

# Start interactive mode
bjarne

# First run will pull the validation container (~500MB)

Then just describe what you want:

> write a function that checks if a number is prime

Generating code...
Validating...
  clang-tidy  PASS (0.8s)
  cppcheck    PASS (0.3s)
  compile     PASS (0.5s)
  asan        PASS (0.2s)
  ubsan       PASS (0.2s)
  msan        PASS (0.3s)
  run         PASS (0.1s)

All validation gates passed!

Commands

Command	Description
/help	Show available commands
/model [haiku|sonnet|opus]	Switch AI model (cost/capability tradeoff)
/save <filename>	Save last generated code to file
/code	Show the last generated code
/validate <file>	Validate an existing file through all gates
/init	Index current workspace for context-aware generation
/config	Show/modify validator settings
/tokens	Show token usage for current session
/debug	Toggle debug mode (logs validation errors to file)
/clear	Clear conversation history
/quit or Ctrl+C	Exit

Configuration

Environment variables:

Variable	Description	Default
BJARNE_PROVIDER	LLM provider: bedrock, anthropic, openai, gemini	bedrock
BJARNE_API_KEY	API key (required for non-Bedrock providers)	-
BJARNE_MODEL	Default model: haiku, sonnet, opus	sonnet
BJARNE_VALIDATOR_IMAGE	Custom validator container image	ghcr.io/3rg0n/bjarne-validator:latest
BJARNE_ASCII	Use ASCII box characters (0 or 1)	1 on macOS
AWS_REGION	AWS region for Bedrock	us-west-2

Model Selection

bjarne uses three model tiers that map to each provider's equivalent:

Tier	Anthropic	OpenAI	Gemini	Best For
haiku	Claude Haiku	GPT-4o-mini	Gemini Flash	Simple functions, quick iterations
sonnet	Claude Sonnet	GPT-4o	Gemini Pro	Most tasks (recommended)
opus	Claude Opus	o1	Gemini Pro	Complex algorithms, architecture

Provider Setup

AWS Bedrock (default):

aws configure  # Set up AWS credentials
# Uses Claude models via Bedrock - no API key needed

Anthropic Direct:

export BJARNE_PROVIDER=anthropic
export BJARNE_API_KEY=sk-ant-...

OpenAI:

export BJARNE_PROVIDER=openai
export BJARNE_API_KEY=sk-...

Google Gemini:

export BJARNE_PROVIDER=gemini
export BJARNE_API_KEY=...

How Validation Works

bjarne runs your code through multiple validation stages in an isolated container:

1. Static Analysis - clang-tidy and cppcheck catch common bugs and style issues
2. Compilation - Strict warnings (-Wall -Wextra -Werror) plus security hardening
3. Runtime Sanitizers - Each catches different bug classes:
   • ASAN: Buffer overflows, use-after-free, double-free
   • UBSAN: Integer overflow, null dereference, alignment issues
   • MSAN: Uninitialized memory reads
   • TSAN: Data races (only when threading detected)

If any stage fails, bjarne sends the error back to the AI with guidance on how to fix it. This loop continues (up to 15 attempts with model escalation) until the code passes all gates.

License

Business Source License 1.1

• Free for: Personal projects, learning, research, academic use, non-profits, internal evaluation
• Commercial production use: Requires a license - contact 62959009+3rg0n@users.noreply.github.com
• Change Date: December 8, 2028 → Apache 2.0

Building from Source

# Clone
git clone https://github.com/3rg0n/bjarne.git
cd bjarne

# Build (standard)
go build -o bjarne .

# Build with ONNX support (enables /init workspace indexing with embeddings)
go build -tags onnx -o bjarne .

Requires Go 1.22+.

Contributing

Contributions welcome! Please read the license terms - contributions are accepted under the same BSL 1.1 license.

---

Named after Bjarne Stroustrup, creator of C++.