OwlSight

On-premise AI-powered code review for CI/CD pipelines.

OwlSight diffs your changes against a base branch, sends them to an OpenAI-compatible LLM with agentic tool-use capabilities, and produces detailed review findings — all without sending code to third-party services.

$ owlsight review --base main --model gpt-4o

  src/UserService.cs
  CRITICAL  SQL injection vulnerability (src/UserService.cs:42-45)
    User input is interpolated directly into SQL query.
    Suggestion: Use parameterized queries instead.

  WARNING   Missing null check (src/UserService.cs:28)
    GetUser() may return null but caller does not check.

  ╭──────────────────────────────╮
  │ Review Summary               │
  ├──────────────┬───────────────┤
  │ Files        │ 4             │
  │ Findings     │ 3             │
  │ Critical     │ 1             │
  │ Warning      │ 1             │
  │ Info         │ 1             │
  ╰──────────────┴───────────────╯

  Review FAILED — critical issues found.

---

Why OwlSight?

Need	OwlSight
Keep code on-premise	Runs against any OpenAI-compatible API — local Ollama, Azure OpenAI, self-hosted vLLM
CI/CD gating	Exit code 1 on critical findings — standard pipeline gate
Context-aware reviews	Agentic loop lets the LLM read files, search code, check git blame
Custom rules	Drop markdown files in .owlsight/rules/ to enforce team standards
No vendor lock-in	Works with any model that supports tool calling

---

Quick Start

Install

# Clone and build
git clone https://github.com/radaiko/OwlSight.git
cd OwlSight
dotnet build

# Or pull the Docker image
docker build -t owlsight .

Initialize

owlsight init

Creates .owlsight/config.json and .owlsight/rules/ with an example rule.

Run a Review

owlsight review --base main --api-key $OPENAI_API_KEY --model gpt-4o

See Getting Started for the full setup walkthrough.

---

Features

AI Code Review

Feature	Description
owlsight review	Review changes against a base branch
owlsight init	Scaffold configuration and example rules
Custom rules	Project-specific review rules in markdown
Agentic loop	LLM investigates code context via tool calls
JSON reports	Machine-readable output for CI integration

LLM Tools

The LLM can call these tools to investigate code context during review:

Tool	Description
read_file	Read entire file contents
read_file_lines	Read specific line range
list_files	List files with optional glob pattern
search_text	Regex search across files
get_file_structure	Directory tree view
get_git_blame	Git blame for authorship info
get_git_log	Recent commit history

See LLM Tools for details.

Output

• Console — Spectre.Console colored output with severity, file:line, title, description, suggestion
• JSON — Structured report with version, timestamp, summary, findings array
• Exit code — 0 = pass, 1 = critical findings, 2 = error

---

Supported LLM Providers

Any OpenAI-compatible API endpoint works. Tested with:

Provider	--base-url
OpenAI	https://api.openai.com/v1 (default)
Azure OpenAI	https://<name>.openai.azure.com/openai/deployments/<deployment>
Ollama	http://localhost:11434/v1
vLLM	http://localhost:8000/v1
LM Studio	http://localhost:1234/v1

---

Requirements

Requirement	Version
.NET	10.0+
Git	Any recent version
LLM API	OpenAI-compatible with tool-calling support

---

Learn More

• Getting Started — Full setup walkthrough
• Configuration — Config file, environment variables, CLI args
• Custom Rules — Writing project-specific review rules
• CI/CD Integration — GitHub Actions, GitLab CI, Azure DevOps
• Docker — Running OwlSight in containers
• CLI Reference — Complete command reference
• Agentic Loop — How the AI review process works
• LLM Tools — Tools available to the LLM during review
• Source Code