Open-source Rust infrastructure for agentic AI
Core Runtime
Agentic AI server. Multi-provider LLM, tool calling, RAG, MCP, multi-tenant metering, OpenAI-compatible API.
Context intelligence engine. Structured business knowledge, workspace isolation, 13 MCP tools.
Code Intelligence
Developer Tools
Self-healing CLI coding agent. 34 tools (AST + LSP + deagle embedded), tiered registry, auto-install. Local llama.cpp / MLX / Ollama + NIM cloud fallback. Compiler-as-auditor.
Declarative config management. TOML manifests, symlink deploy, drift detection (aegis watch), encrypted secrets, overlay networking, toolchain orchestration.
Minimal agent runtime for direct NIM/NVIDIA inference. Zero built-in tools — consumers register their own. 10 deps. Single binary.
Self-contained web search MCP server. 7 backends with automatic fallback. Works from any IP.
MCP server for Eruka context memory. 13 tools, 4 knowledge states, anti-hallucination constraint injection.
Production WASM build tool for Leptos frontends. Fixes wasm-opt bulk-memory, content hashing, index.html patching.
NVIDIA NIM model latency benchmarker. Ping, tool-use, and agent task timing across all NIM models.
Platform
Execution context engineering. Unified tool abstraction, skill workflows, session management.
Skill distillation. Teacher-student LLM distillation for SKILL.md generation, evaluation, and refinement.
Development stack for AI-assisted multi-repo work. Persistent memory, quality gates, deployment automation.
End-to-end llama.cpp toolkit. API client, HuggingFace Hub downloads, server orchestration, and a 5-test benchmark suite for comparing model quantizations.
Component library for Leptos frontends. 29 accessible, signal-driven components. On crates.io as dui-leptos.
Case Studies
Crisis response requires architectural guarantees, not prompt engineering. Eruka's UNKNOWN state blocks the agent from inventing hotline numbers or clinical advice. Constraint injection enforces scope at runtime.
Eruka's four knowledge states map 1:1 to student mastery: UNKNOWN → UNCERTAIN → INFERRED → CONFIRMED. The agent knows what each student has learned, forgotten, and relearned.
DIRMACS runs on the dirmacs stack. Every repo, every session, every commit goes through ARES + Eruka + Pawan. The platform is built by its own platform.
Our circular methodology for taking any operation from manual to autonomous. DIRMACS is its own first client.