An "Executive Function as a Service" self-iterating AI coding agent that conducts internal "Society of Mind" deliberation before execution and autonomously loops until code quality meets the threshold. Built in Rust as a single static binary with zero runtime dependencies.
Core Positioning#
OpenKoi addresses the lack of autonomous review and iteration in traditional AI coding tools by introducing a "parliamentary" multi-dimensional deliberation mechanism through the Plan-Execute-Evaluate-Refine loop for automated quality control, focusing on end-to-end autonomous execution of code refactoring, review, and multi-step programming tasks.
Society of Mind Architecture#
Layered architecture: top-level Sovereign Directive (SOUL.md + Value Model + Trajectory) passes down to Parliament (5 Agencies: Guardian, Economist, Empath, Scholar, Strategist), then Orchestrator dispatches to Executor, Evaluator, Learner, Pattern Miner, and Integrations.
Autonomous Iteration#
- Plan-Execute-Evaluate-Refine loop where the agent is its own reviewer
- Quality threshold self-check (default quality ≥ 0.8) with automatic redo
--simulatemode for dry-run simulation
Memory & World Model#
- Persistent memory: SQLite + vector search, cross-session persistence
- World Model: Tool Atlas (tool reliability/failure modes), Domain Atlas (domain knowledge), Human Atlas (user preferences)
- Pattern mining: observes user habits, auto-suggests new Skills
Cognition & Reflection#
- Daily/weekly/deep reflection loops, cognitive honesty audits
- Soul personality system that evolves with interaction patterns
- Introspection via
openkoi soul/mind/reflectcommands
Multi-Model Orchestration#
8+ providers (Anthropic, OpenAI, Google, Ollama, AWS Bedrock, Groq, DeepSeek, Moonshot/Kimi, MiniMax, OpenRouter, Together, xAI, Qwen, etc.), dynamic model discovery with fuzzy matching, role-based model assignment (executor/evaluator/planner/embedder can use different models).
Security & Permissions#
- Auto-redaction of sensitive info (keys, passwords, PII) before sending to AI
- Trust & delegation system: per-domain autonomy levels (ask/suggest/act/autonomous), auditable
Extensions & Integrations#
- 3-layer plugin system: MCP (external tools), WASM (wasmtime 41 Cranelift JIT sandbox), Rhai (scripting engine)
- 10 integrations: Slack, Discord, MS Teams, GitHub, Jira, Linear, Notion, Google Docs, Telegram, Email
- OpenClaw-compatible
.SKILL.mdformat, portable across tools
Operations & Monitoring#
- Real-time cost tracking with budget Webhook alerts
- TUI real-time dashboard (ratatui)
- Local REST API (axum, port 9742) with Webhook callbacks
Performance#
- <10ms startup, ~5MB idle memory, ~20MB binary size
- Platforms: Linux (x86_64, ARM64), macOS (Intel, Apple Silicon)
Installation#
brew tap openkoi-ai/openkoi && brew install openkoi
# or
cargo install openkoi
# or
curl -fsSL https://openkoi.ai/install.sh | sh
Zero-config startup — auto-detects credentials from env vars, existing CLI tools (Claude CLI, Qwen CLI), macOS Keychain, and local model servers.
Key Commands#
| Command | Description |
|---|---|
openkoi "task" | Run task (default 3 iterations) |
openkoi think "task" | Full EFaaS pipeline |
openkoi think "task" --simulate | Simulation mode |
openkoi chat / --resume <id> | Interactive REPL / resume session |
openkoi status --live | Live monitoring |
openkoi dashboard | TUI dashboard |
openkoi connect copilot | Connect GitHub Copilot subscription |
Key params: -i N (max iterations), --quality F (quality threshold), --quiet, --redact, -m MODEL.
Open Source#
MIT (Copyright © 2025 OpenKoi)
Unconfirmed Information#
- Team members/authors: no individual developer info listed
- First release date: not explicitly stated
- openkoi.ai website content not verified by actual visit