// the find
HKUDS/DeepCode
"DeepCode: Open Agentic Coding (Agent Harness & Loop Engineering & Multi-Agent Orchestration)"
DeepCode is a general-purpose coding agent (TUI, Desktop, and Web front ends over one shared local service) built by HKU's Data Intelligence Lab. It started life as Paper2Code — turning research papers into working implementations — and has since pivoted into a Claude Code / Codex-style agent harness with goal-driven loops, MCP support, and subagent delegation. Best fit for developers who want an open-source, self-hosted alternative to those tools and don't mind a fast-moving, sprawling codebase.
The goal loop (implement, verify, repair) keeps an evidence ledger and can resume across restarts or model switches instead of treating one plausible answer as done. Permissions are three-valued with real sandboxing (Windows gets its own Job Object isolation), and MCP servers can be loaded lazily via deferLoading with OAuth handled through a loopback callback rather than a browser popup at startup. Subagent support goes further than most competitors — it can shell out to installed Codex or Claude Code CLIs with a scrubbed environment, not just spawn native subagents.
The command-safety story has needed repeated patching — the changelog shows at least three separate fixes to the dangerous-command screen (case/whitespace bypass, argv-vs-substring matching, a native mkdir/rm/cp fast path skipping the sandbox entirely), which is a bad sign if you're trusting it to gate destructive shell commands. The surface area is enormous (app_server, core/harness, core/mcp, core/plugins, providers, dictation, automation...) for what's fundamentally a CLI coding assistant, meaning a lot to audit before you'd trust it on a real repo. Session state lives in a custom SQLite layer with its own compaction and locking logic rather than reusing a proven session format, and the README leans hard on badges and marketing copy over benchmarks against the tools it's actually competing with.