Local-First Forensics
Built for Developers
Who Treat AI like Software
AI agents fail in highly complex, non-deterministic ways. I built Orchid to make AI behavior as debuggable and testable as normal code.
The Story
In late 2025, I was building multi-agent systems. Whenever an output went off the rails, or a validation schema failed, I spent hours digging through JSON logs and trace streams. Existing tools required intrusive and time-consuming code instrumentation, littering my codebase with custom decorators, client overrides, and telemetry boilerplate that felt entirely decoupled from clean application logic.
I realized the solution was simple: use an intercepting proxy. By routing local traffic through a lightweight proxy engine, I could capture inputs, outputs, and token counts without intrusive SDK lines or AST rewriting.
Orchid was born. It separates payload recording from runtime execution, storing trace runs locally in SQLite for offline playback. Traces can be explored dynamically through the web visualizer, or queried programmatically using the built-in MCP server.
No intrusive code instrumentation. No sending your data to a third party.
First Principles
Zero-Intrusion Telemetry
Across all supported languages, Orchid hooks into your existing HTTP transport layers rather than requiring you to adopt proprietary AI client wrappers.
Local-First SQLite Cache
All traces persist locally in SQLite. Telemetry remains completely in your control, operating with zero latency and 100% security boundary integrity.
Deterministic Playback & Test
Export JSON fixtures, override sessions, and run offline playback runs in CI/CD with zero API cost using native test framework integrations.
Web UI & MCP Integration
Inspect runs visually through the Web UI, or integrate the MCP server directly into your IDE agent to query trace details.