TricorderAgent 1.0 is out — open source & local-first

Software & AI Creative

Software developer from Austria building local-first AI — a personal assistant, a generative video studio, an open-source agent, and custom fine-tuned models, all running on my own hardware.

Explore the products Get in touch

3 AI Products

40 Agent Tools

32GB VRAM

One machine in Vienna. No cloud. Everything below runs here.

What I do

Services

All AI runs locally on dedicated hardware — no cloud, full privacy.

New · Featured

Custom LoRA Fine-Tuning

Your model, trained on your data — fully private, on local hardware. LLM assistants fine-tuned from real conversations (Qwen3.8 27B), style & character LoRAs for Qwen-Image and FLUX.2, video LoRAs for WAN 2.2, LTX-2.5 and MiniMax H3.

The full pipeline, not just a training run: dataset engineering from your raw data → QLoRA training → automated A/B evaluation against your current model → delivery as GGUF or safetensors with model card and checksums.

Latest build: a 27B assistant fine-tuned on 500+ curated real conversations with full tool-calling — dataset, training, quantization and a 36-case evaluation suite, end to end on a single RTX 5090.

QLoRAQwen3.8 27BWAN 2.2LTX-2.5MiniMax H3Qwen ImageFLUX.2

AI Video Generation

Clips up to 60s with synchronized native audio — WAN 2.2, LTX-2.5 and MiniMax H3 on a local RTX 5090, upscalable to 4K.

ComfyUIWAN 2.2LTX-2.5MiniMax H34K

Web Development

React, Vite, TailwindCSS — modern web apps and PWAs with real-time data, auth, offline.

ReactTypeScriptPWAVite

AI Image & Audio

4K generation via ComfyUI — Qwen Image, inpainting, style transfer, face restore, voice synthesis.

Qwen ImageStable DiffusionVoice AI

AI Setup & Consulting

ComfyUI & llama.cpp configuration, local agent setups, fine-tuning pipelines, dataset engineering, hardware optimization.

ComfyUIllama.cppAgentsFine-tuning

Local Hardware

RTX 5090 32GB Core i9 10C/20T 64GB RAM 2×1TB NVMe 6TB Backup Ubuntu Linux

Model Stack

Qwen3.8 27B LLM WAN 2.2 Video LTX-2.5 22B Video+Audio MiniMax H3 Video+Audio Qwen Image Stills Whisper · XTTS Voice

Not just services — the same tools I build and rely on every day.

Flagship products

Tricorder & Motionstill

Two standout products, one machine — a personal AI assistant and a generative video studio, both running fully local on a dedicated RTX 5090 workstation.

A fully local AI assistant in a Star Trek LCARS shell, powered by Qwen3.8 27B on the RTX 5090 — now being fine-tuned on its own real conversations. No cloud, no filters, no compromises: raw streaming inference with speculative decoding, an in-app agent loop with native function calling, and a toolbelt of 115 native tools loaded on demand — from browser automation and a Docker code sandbox to rendering images on ComfyUI.

126 T/s Peak Inference

115 Tools Native Toolbelt

100% Uncensored

Qwen3.8 27B · custom fine-tune Hybrid linear-attention model on llama.cpp with MTP speculative decoding — LoRA-tuned on the assistant's own chat history

Agent mode & 115 native tools Autonomous multi-step tasks — browser, code sandbox, git, ffmpeg, ComfyUI rendering

Self-improving skills & memory Distills reusable skills from its own chats; persistent memory with semantic vector recall

Voice, vision & sensors Speech in and out, camera scanner with AI analysis, live sensor readouts — plus WhatsApp, Spotify & health

Qwen3.8 27Bllama.cppCustom LoRALCARS UIAgentVector MemoryPWARTX 5090

A local studio that turns the workstation into a full generative pipeline. The browser talks straight to ComfyUI: it builds the graph, queues the render and streams the sampler's live preview back while it works. Twelve modes span stills, video, instruction-based editing, synchronized audio+video and reference-guided generation — and nothing ever leaves the machine.

12 Modes Generation

60 s Max Clip

~2 s Turbo Render

Video with native audio LTX-2.5 and MiniMax H3 render voice, sound effects and music in the same pass as the video

Reference→Video Up to 9 reference images plus clips and audio lock identity, wardrobe, style and scene

Directed like a film Multi-shot cuts, LLM prompt enhancement, voice cloning and auto duration prediction

Edit, restore, upscale Instruction-based image & video edits, face restore, 4× upscaling and frame interpolation

ComfyUIWan 2.2Qwen ImageLTX-2.5MiniMax H3RTX 5090

Motionstill studio interface — mode cards for Text to Image, Text to Video, Image to Video and Image Edit above prompt fields and quality settings

Also built

More projects

Live

MotionStill: Cut

A film editor for prompts — DaVinci-style NLE where a local LLM writes the shots and ComfyUI renders them.

AI VideoNLEMiniMax H3LTX-2.5

PixelStill

Local AI image editor — Photoshop-style layers, masks & adjustment layers, backed by ComfyUI for generate, inpaint, expand and upscale.

AI ImageLayer EditorComfyUI

Under Review

Nexus Monitor

AI workstation monitor & control panel with real-time GPU tracking.

AIDashboardReal-time

Safety Analytics

Enterprise safety platform — 33,641 lines, 4-tier privacy, PDF export.

EnterpriseAnalyticsGDPR

Safety Hub

PWA for safety auditing — React 18, TypeScript, Auth0, real-time analytics.

ReactPWAAuth0

Safety Quest

16-bit retro game — 12 chiptune tracks, 9 boss battles, 100+ hazard scenarios.

GameCanvasWeb Audio

Healthy Brunch Club

Restaurant PWA — Vue.js, headless CMS, serverless functions.

Vue.jsCMSServerless

New · Public release · v1.0

TricorderAgent

A local-first AI agent that actually does things.

The agent engine behind Tricorder — stripped down, open-sourced, MIT licensed. It doesn't just chat: it reads and writes files, runs commands, searches the web, executes code in a sandbox, drives a browser, and remembers what matters across conversations. Point it at a model you run yourself — LM Studio, llama.cpp, or Ollama — and it works. No account, no API key, no data leaving your machine.

~100 Tokens / sec

40 Tools · 8 categories

256K Context window

0 npm dependencies

TricorderAgent mid-task: three tool calls with arguments and timings — grep and read_file done, write_file running — and an approval card asking permission before the file is written

Mid-task: every tool call visible as it happens, with an approval card before anything is written — streaming at 103 tokens per second.

Real tools, not just chat

Files, shell, web, git, SQLite, PDF/DOCX/XLSX extraction — 40 tools it can actually call, with live streaming output.

Sandboxed code execution

Python, Node and Bash run in throwaway, network-isolated Docker containers — results come back, side effects don't.

Web & browser automation

Searches, scrapes and cross-checks sources — and drives headless Chromium: navigate, click, type, read the page visually.

Memory that persists

Remembers facts about you and your projects across conversations, ranked by relevance and injected into its own prompt.

Scheduling & sub-agents

Cron-style tasks and reminders that run without you, plus background agents for long work — results collected when done.

You stay in control

It plans before it acts, asks before it writes files or runs commands, and shows every tool call as it happens.

terminal

$ git clone https://github.com/eres2k/TricorderAgent.git
$ cd TricorderAgent
$ npm start

That's the whole install. First run finds your model server, sizes the model to your memory and configures itself — zero npm dependencies, Node's standard library only.

Not ready to install anything? The browser build runs the same interface as a plain web page: point it at your own LM Studio, llama.cpp or Ollama and it works, with an in-browser workspace standing in for the tools that need a real machine. Nothing is sent to this site — there is no backend behind that page.

Run it in your browser View on GitHub Read the overview MIT licensed · LM Studio, llama.cpp & Ollama

And this is what comes out of it.

Creative work

Visuals & Music

AI-generated effects, cinematic clips, and remastered covers — all produced locally with the same pipeline that powers Motionstill.

Short-form series in production — everything below is filmed on this machine, fully local.

In production No cloud. Watch. One GPU renders video + audio, live

In production Reference → Video 9 images lock identity, wardrobe & style

In production Text → talking video Voice, sound & picture in a single pass

In production 2-second renders Five prompts, five images, real time

In production An agent that does things Email to calendar, hands-free, on-device

In production I trained my AI on itself 740 chats → a custom 27B fine-tune

In production Style LoRA in 30 minutes 20 images in, before/after grid out

In production Base vs fine-tune Same prompt, scored A/B — honest results

Select a track Browse below

0:00 0:00

About

Skills & Tech

Frontend

JavaScriptTypeScriptReactVue.jsHTML5CSS3TailwindCSS

Tools & Runtime

Node.jsViteChart.jsPWAService WorkersWeb Audio APICanvas API

AI & Agents

LLM AgentsTool Callingllama.cppComfyUIAI VideoLocal Pipelines

Model Tuning

LLM Fine-tuning (QLoRA)Diffusion LoRAsVideo-Model LoRAsDataset EngineeringGGUF QuantizationA/B Model EvaluationUnslothComfyUI Model Integration

Domain

Video EditingVFXAudio ProductionWorkplace SafetyGDPRData Analytics

Get in touch

Let's work together

Need a model tuned on your data, a private AI pipeline on your own hardware, or generative video at 4K? I've built all three for myself — I can build them for you.

GitHub@eres2k

LinkedInErwin Esener

Emailerwin.esener@gmail.com