AIMindMesh is a privacy-first, local-first agentic ecosystem designed to orchestrate a mesh of heterogeneous nodes—Android devices, high-performance PC clients, and specialized server orchestrators—into a single, unified cognitive fabric. It is the first assistant capable of autonomous self-evolution, repairing and improving its own source code through a multi-path agentic loop.
AIMindMesh breaks the traditional "Wrapper" paradigm by implementing a strict Native/Frontend Split:
- Computational Bedrock (Native C++/Kotlin/Rust): Executes heavy-lift operations—LLM inference, NPU delegation, high-fidelity audio processing, and vector similarity—directly on bare metal.
- State Orchestration (React/TypeScript): Manages complex business logic, UI state, and real-time synchronization via a custom Robust Proxy Pattern.
- Distributed Mesh Topology: Nodes (Mobile, PC, VPS) communicate via a secured WireHole (WireGuard) tunnel, sharing inference power, knowledge deltas, and task execution.
The mobile node is not just a client; it's a high-performance inference powerhouse optimized for modern Snapdragon silicon that could also run standalone.
- Adreno OpenCL & Vulkan: Optimized GGUF inference using the Qualcomm-contributed OpenCL backend for Adreno GPUs, significantly outperforming generic Vulkan implementations.
- Hexagon NPU Delegation: Direct routing of LiteRT (
.litertlm) models to the Qualcomm Hexagon HTP via QNN delegates, achieving studio-level tokens-per-second with minimal thermal impact. - Speculative Decoding (MTP): Implements Multi-Token Prediction (LiteRT 0.11.0) to predict and verify multiple tokens per forward pass.
- Persistent KV Cache: Disk-based serialization of conversation states to
cache/litert_cache/, allowing instant resumption of context after app restarts or backgrounding. - VRAM Guardian: Dynamic RAM-pressure scaling that monitors
onTrimMemoryto prevent OOM kills by proactively compressing history via local summarization.
- Offline TTS Engine (Kokoro & Piper): High-quality, real-time voice synthesis running entirely on-device without cloud dependencies.
- Kokoro TTS (v1.0): State-of-the-art multi-lingual voice model integrated via a statically-linked Sherpa-ONNX backend. Features dynamic runtime language switching (e.g. English, Italian) and seamless in-app bundle downloading or local
.tar.bz2importing. - Piper TTS: Fast, private synthesis using raw ONNX models with dynamic voice management.
- Kokoro TTS (v1.0): State-of-the-art multi-lingual voice model integrated via a statically-linked Sherpa-ONNX backend. Features dynamic runtime language switching (e.g. English, Italian) and seamless in-app bundle downloading or local
- 3-Pass Diarization:
- Profiling: Global clustering of ECAPA-TDNN voice embeddings.
- Classification: Segment assignment based on speaker centroid proximity.
- HMM Smoothing: Viterbi decoding to eliminate spurious speaker oscillations.
- Voxtral Realtime: 4B Multimodal STT (PCM → mel → CLIP → llama) for near-zero latency voice interaction.
- Durable Recording: Direct-to-disk PCM encoding ensures reliability for long sessions (>3h) without memory exhaustion.
- Driving-Optimized UI: Custom
GridTemplatedashboard for safe, hands-free interaction. - Seamless Sync: Real-time access to your Agenda, Kanban, and Assistant Call directly from the car's head unit.
- Privacy-First Call Mode: Routes audio through the earpiece or car speakers with full VAD-based turn-taking.
The server acts as the "Central Nervous System", managing long-term memory and the ecosystem's autonomous growth.
- Tiered Task Prioritization: Dynamically routes tasks (Embeddings → Lightweight → Complex → Evolution) across the mesh based on node hardware capability, proximity, and quota. The routes are customizable via Client PC and Admin Mobile per task type (e.g. embed, lightweight, complex, evolution).
- Claude First, While the Credit Lasts: the monthly Claude API credit of the Max plan answers chats, research, debates, wiki pages and code generation first (Haiku, Sonnet or Opus per category, prompt caching, half-price Message Batches for the scheduled cycles), and the server agents (OpenClaw, Hermes, venture builds) through the same proxy; when the credit or the internal budget runs out, the usual chain (Gemini, OpenRouter, FreeLLMAPI, local nodes) takes over.
- Neural Wiki & Knowledge Graph (Neo4j): Automatically synthesizes raw meeting data and memories into a structured Neo4j knowledge graph, creating a persistent, searchable "Neural Wiki" of your entire digital life.
- FCM Proactive Push: Real-time delivery of server-generated "Neural Insights" directly to mobile notification trays via Firebase.
Orchestrates autonomous operations, team roles expansion, and automated verification:
- AI Council Debate: A short multi-role review that ends with structured votes; a two-thirds majority approves, and only a human approval starts a build.
- Gitea VCS Provisioning: Repository creation with the generated technical specification, done by the server.
- Venture Build Loop: The OpenClaw agent works without credentials in a server-managed folder; the server snapshots each run, validates it in a throw-away Docker container (install online, build and tests offline) and pushes it only if it passes. Every loop has a budget and a kill switch.
- HR Recruitment Service: Analysis of the organization's gaps, proposing new Agent roles that a human materializes.
- Meeting-to-Organization Bridge: Scans meeting transcripts on mobile nodes to extract goal/directive candidates, selectively routing them to the AI Council.
AIMindMesh actively improves its own source tree through three distinct generation paths:
- Server-Native Evolution: Orchestrates multi-file contexts to ensure architecturally sound patches, delegating to Claude (while its credit lasts), then Gemini or Openrouter for complex refactoring.
- Agentic OpenClaw Loop: High-autonomy worker for user-requested projects that need external research.
- On-Device Termux Scripting: Local models generate and execute bash/python scripts via the native Termux Bridge for system-level Android optimizations.
A lightweight Tauri-based client that bridges high-performance PC hardware into the mesh.
- Ollama Bridge: Lends local GPU power to mobile nodes for complex reasoning.
- Telemetry Dashboard: Real-time monitoring of CPU/RAM/Thermal states across the entire mesh.
- Admin Mobile (
aimindmesh-admin-mobile): Capacitor app to operate the server from the phone — inference queue, nodes, logs, Kasm sessions, document ingestion, auto-evolution proposals and server configuration (secrets are shown masked). - Hermes (
aimindmesh-hermes): NousResearch Hermes agent container, an alternative agent runtime next to OpenClaw, configured from the*.examplefiles in its folder.
- WireHole VPN: WireGuard tunnel + PiHole (DNS privacy) + Unbound (Recursive DNS).
- Gitea: Self-hosted Git service for autonomous VCS and evolution patches.
- SearXNG: Private metasearch engine ensuring untracked web research.
- FreeLLMAPI: Private, self-hosted LLM gateway ensuring OpenAI-compatible model routing and credential isolation. (https://github.com/tashfeenahmed/freellmapi)
- Anthropic Claude API: the plan's monthly credit, called by the server with its own key (never given to the agents); spend, budget and cache hits are visible in the admin apps.
- Kasm Workspaces: Secure, containerized environments for agentic execution and shadow testing.
- Neo4j Graph Database: The containerized knowledge engine powering the "Neural Wiki."
- Foundation: Run
./deploy_infrastructure.shon your VPS. This sets up the networking, Gitea, SearXNG, and Kasm, while automatically disabling any native Neo4j services to prevent port conflicts. - Brain: Deploy
aimindmesh-serverusing Docker Compose. Use./deploy_to_cloud.sh --fullfor a complete automated setup of the Server, Neo4j, and OpenClaw gateway.- Create
config.jsonfromconfig.json.exampleandsrc/config.tsfromsrc/config.ts.example. Set a strongserver.apiKeyandautoEvolution.giteaWebhookSecret(use the same secret in the Gitea webhook settings). - Place
firebase-service-account.jsonnext todocker-compose.yml: it is mounted at runtime, never baked into the image. - The server (3030) and Neo4j (7474/7687) are published only on
127.0.0.1and the VPN address10.2.0.1.
- Create
- Nodes: Point the Mobile app, the PC client and the Admin mobile app to
http://10.2.0.1:3030over WireGuard, with the server API key (sent asx-api-key, and as?token=on WebSockets). - CI (optional): register a Gitea Actions runner with the hardened example in
infrastructure/docker/services/gitea/act_runner/to run.gitea/workflows/ci.yml(seeDEVELOPMENT_GUIDE.md).
The OpenClaw agent requires a configuration folder containing its skills and auth tokens.
- Copy the template folder:
cp -r aimindmesh-server/openclaw-config.template aimindmesh-server/openclaw-config - Open
aimindmesh-server/openclaw-config/openclaw.jsonand insert your Telegram Bot Token and define a Gateway Auth Token. - The folder is automatically ignored by Git to protect your tokens.
The repository includes *.sh.example templates for rapid deployment (e.g., deploy_to_cloud.sh.example, publish_android.sh.example). Copy each one without the .example suffix, configure, and run.
- Use
./deploy_to_cloud.sh --fullto deploy and configure the entire server stack (Server, Neo4j, OpenClaw). - These files are git-ignored to protect your private credentials.
Licensed under the PolyForm Noncommercial License 1.0.0.
- Free for personal, educational, and research use.
- Commercial use requires a separate, paid license agreement.
Architect & Designer: Andre (@aimindmesh)
Development Support: Co-authored and implemented in collaboration with Gemini & Claude.
Contact: aimindmesh@proton.me
Philosophy: Privacy is a right, Autonomy is the goal. Designed by Human intelligence, evolved with Artificial Intelligence.