Agents with personality
Each agent is six Markdown files, editable from the interface itself: soul, skills, tools, memory, heartbeat and profile.
HydraOps is a multi-agent system that runs on your machine. Agents with their own personality write code, answer questions and generate images and video, with API or local models. Your keys never leave home: everything goes through a credential firewall.
I use it every day on my own machine. There is a Windows desktop installer and a headless server mode; a container image is on the way.
Real recordings of the app with real models. The waits while the model thinks are sped up.
If you have tried OpenClaw-style personal assistants, this is the multi-agent take: several agents in one chat, orchestrated locally. No mandatory external services. Your keys and your data never leave your machine.
Each agent is six Markdown files, editable from the interface itself: soul, skills, tools, memory, heartbeat and profile.
Code, general, image and video. Each one has its own engine and resolution, and you can configure them per agent.
Native add-ons, your own in my_addons/ with hot loading, and MCP servers over HTTP.
API keys are never in the repository, the database or the .env: a local proxy injects them at the network boundary. Workers only ever see the "proxy" placeholder.
Every tool goes through a filter that blocks credential paths, catastrophic commands and requests to internal networks, and redacts secrets from results.
Attachments, inline images and video, scheduled tasks (cron), statistics and an interface in five languages: es, en, it, fr and pt-BR.
Apps on your computer and online services on top, HydraOps' own tools below. Tokens stay in the local key proxy.
Builds and edits scenes in the Blender you have open.
Read the post →Runs the workflows you saved: images, video and 3D on your own GPU.
Read the post →Builds scenes in your Unity 6 editor, runs tests and builds.
Microsoft's real browser: pages that need JavaScript, clicks, forms, screenshots.
Microsoft's converter: PDFs, Office files and web pages into Markdown.
Web results from the Brave Search API, by region.
One up-to-date answer with its sources, instead of a list of links.
Sign in with your ChatGPT plan and the agents use it, no API key.
Local models on your machine through Ollama, llama.cpp, LM Studio or vLLM.
Reads code, triages issues and PRs, comments. Your token decides what it may write.
Read the post →Talk to your agents from the phone and approve held actions there.
Read the post →Posts to your accounts, several if you like. Every post waits for your approval.
Searches models and datasets, reads licenses, checks gated access.
Finds checkpoints and LoRAs with their trigger words. Read only.
A shared calendar with notes and tasks scheduled for an agent. Changes wait for your approval.
Runs a prompt by itself on a schedule; the result arrives in the chat or on Telegram.
Turns a description into a 3D model with your own model, rendered in a sandbox.
Read the post →Keeps short notes that apply to every future conversation.
Read the post →Searches your past conversations by keyword.
Read the post →Picks a ComfyUI workflow you saved, changes only what you asked and brings back the result.
Read the post →/plan proposes the steps. Nothing runs until you approve.
Read the post →Draw the task as a diagram and get one self-contained prompt.
Read the post →Know-how for the agents from the catalog or any GitHub repository, checked before it installs.
Type / for verbs that cost no tokens: /agents, /status, /remember. Also on Telegram.
Hands a task to another agent; the reply lands in that agent's chat.
Sends a message or a result to your phone when you ask for it.
Crops, rotates, resizes, trims the background or puts several images on one sheet.
Searches the web and returns a short list of links. No key needed.
Reads a page you already know as Markdown. Also reads RSS and Atom feeds.
Reads a video's transcript to summarize or quote it.
HydraOps is not affiliated with or endorsed by the companies named here. All product names, logos and brands are property of their respective owners.
A pnpm monorepo in TypeScript ESM. No service publishes directly to NATS: everything is written to a SQLite outbox first and a single process drains it, so a network failure never loses events.
Four things already taken care of, so you don’t have to think about them.
They live outside the project and the key-proxy injects them at the network boundary: neither the workers, nor the database, nor the .env ever see a real key.
It blocks credential paths and destructive commands, redacts secrets from results, and keeps fetch_url away from internal addresses.
Out of the box it is not reachable from another machine. Nothing gets exposed by accident.
With HYDRA_HOST=0.0.0.0 the API demands a HYDRA_AUTH_TOKEN: the browser asks for it once, and with no token defined the API simply refuses to open up.
To report a vulnerability: security@hydraops.org
Three ways to get it running, from fastest to most flexible.
Download the .exe from Releases. It needs no Node and no pnpm: the services run on the Node bundled with Electron. Your data and keys stay outside the installation directory and the app updates itself.
Download .exeFor Windows, Linux and macOS. Requirements: Node 20+, pnpm 9 and the nats-server binary (on your PATH, in a nats/ folder inside the repo, or via NATS_SERVER_BIN in the .env). No build steps to remember: pnpm desktop builds packages and interface and opens the app window.
git clone https://github.com/TraX22/HydraOps.git
cd HydraOps
pnpm install
cp .env.example .env
pnpm desktopFor a screenless mini PC: a single command builds, brings up NATS and all eight services, applies migrations and seeds the database, health-checks each phase and restarts anything that crashes. A systemd unit ships in deploy/.
pnpm serve