Cordon - Sandboxed AI Agent Harness
A provider-neutral agent harness, inspired by Hermes Agent and rebuilt where its design did not fit: about 12,000 lines of Python with around 1,000 automated tests.
- Built an agent loop for Anthropic, OpenAI-compatible and local llama.cpp models, with parallel tool calls, parallel subagent delegation and an automatic model fallback chain.
- Designed security controls for autonomous agents: a sandbox that fails closed, network egress through an allowlist proxy, a scrubbed environment, and an approval gate on every tool call.
- Implemented a tamper-evident, hash-chained audit log with documented SIEM signals for egress, approvals, configuration changes and failures.
- Built a hardened local web UI (token auth, Host and Origin checks, strict CSP) with streaming chat, a live view of running subagents, voice control, and context compaction for long sessions.
- Python
- Anthropic and OpenAI APIs
- llama.cpp
- Docker
- macOS sandbox (seatbelt)
- SQLite
- pytest
- JavaScript