Home

Personal project

Argus

Argus is the agent system I built on top of Odysseus. Its assistant, Jarvis, runs on a model on my laptop, keeps notes he can search and talks by voice. The name comes from the hundred-eyed watchman of Greek myth, who never slept.

Built on Odysseus, the open-source self-hosted AI hub (AGPL-3.0).

Evaluations, local model13/13Qwen3.6-35B-A3B, thinking off, median 2.6 s
Voice, first word0.7 sMid-conversation, on the laptop's GPU
Memory server7 toolsOver MCP, 29 tests

How it fits

capability level

My layer on top of Odysseus

Argus, my work Odysseus other projectsswipe →
Three ways in (voice line, hub chat, Claude Code) reach the Odysseus hub, where Jarvis runs as a persona. Argus plugs into the hub: a memory server over MCP backed by a plain-text vault, a bridge to Claude Code, a bridge to free cloud models, a bridge that brings Windows tools such as REA and Scrapling into the hub, and a speech service with the Daniel voice and Whisper. Underneath, one local model server runs Qwen3.6-35B-A3B on the laptop's GPU, shared by Jarvis, Hermes and the voice line, which talks to it directly. Voice linewake word, local brain Hub chatin the browser Claude Codein the terminal Odysseus hub chat · agent loop · MCP client · personas · skills · model endpoints · tool approvals Jarvis persona Jarvis skills Voice controls re-shown jarvis-memoryMCP · 7 tools claude-bridgehard jobs free-bridgefree fallback tool bridgeWindows tools in jarvis-voiceDaniel · Whisper Vault, plain text Claude Code Free cloud models REA · Scrapling Hermes Agentagent, and the model server llama.cpp · Qwen3.6-35B-A3B, one shared copy my Vulkan fix: snapshot 0.61 s to 0.03 s

Jarvis is an Odysseus persona. His memory, the other models and the outside tools all reach him through my layer. The voice line skips the hub and talks to the model directly.

Part of Work · What I'm doing

I wanted an assistant that keeps my notes on my own machine and answers routine questions in seconds. Argus is that system, built as a layer on Odysseus: a memory server, bridges to Claude and to free cloud models, a voice line and a set of evaluations. With thinking switched off, the local model passes all 13 evaluation tasks, with a median answer time of 2.6 s.

The hub below is a redrawing in my colours with sample conversations. Press Next for a tour, or click around.

Terms used here
Agent
A language model working in a loop: it calls a tool, reads the result and decides the next step.
Local model
A language model running on my laptop instead of in the cloud. Here, Qwen3.6-35B-A3B on the laptop's GPU.
MCP
Model Context Protocol, an open standard for connecting models to tools. A server lists its tools and the model calls them.
Evaluation (eval)
A fixed set of tasks with automatic checks, run again after a change to see whether the system still works.
Persona
The identity and rules a model is given. Jarvis is an Odysseus persona.
Routing
Choosing which model gets a job: local for routine work, Claude for hard multi-step jobs, free cloud models as a fallback.
Thinking mode
A setting where the model reasons at length before answering. Slower, and here it twice used up the whole answer budget.
Trace
A timed record of each step of a request, for finding what is slow or broken.
Wake word
"Jarvis", the word that tells the voice line to start listening.

The hub, recreated

Tour · 6 stopsPress Next, or click aroundThe hub, redrawn in my colours, with sample conversations.

Jarvis
Local · Qwen3.6Claude · bridgeFree cloud

Explain in two lines why you run on a local model.

Jarvis · Qwen3.6-35B-A3B · local

Your notes stay on this machine, and routine answers come back in seconds. When a job needs more, I hand it to Claude and tell you I did.

    Try it

    runs in your browser · nothing is sent

    Which brain would Argus use?

    Each request goes to one of three places: the local model, a free cloud model or Claude. A small learned router makes that choice before any model runs, and personal data never goes to a free model. This is the same router, running in your browser.

    Why a separate router. I tested the local model on 16 routing tasks it had never seen. Left to route itself, it passed 7. With the router narrowing the tools it is offered, it passed 15; on the last one the router picked the free cloud model and the local model answered anyway. Writing the same advice into its instructions, with the first version of the router, passed 8. A small model follows a narrower choice more reliably than a hint.

    The router is a logistic regression over word and character patterns, 175 KB in this page. I trained it on 160 examples I wrote by hand plus 1,962 written by Claude Sonnet 5.5, each confirmed by Claude Opus 5.5 labelling blind. On 394 requests from topics it never saw it is right 98% of the time, against 71% with my examples alone, and it routes all 16 test tasks correctly. In Argus (Python) a decision takes about 40 ms, and a personal-data screen for emails, phone, card and ID numbers, addresses, health and money can override it. That screen is not on this page.

    Ownership

    who built what

    What I built and what Odysseus provides

    Argus · mine

    • Jarvis in the hub persona, skills and an import builder
    • jarvis-memory vault memory over MCP, 7 tools, 29 tests
    • claude-bridge hard jobs to Claude Code, read-only by default, never a shell
    • free-bridge free cloud models as a fallback, text only
    • Tool bridge brings Windows-only MCP tools into the hub
    • Voice the local brain and wake word for the voice line; Daniel's voice inside Odysseus
    • Routing local model for routine work, Claude for hard multi-step jobs, free cloud as fallback
    • Measurement 13-task evaluations, traces, a performance dashboard, a system audit
    • Supervisor a tray app that runs the services and restarts any that crash
    • Engine fix llama.cpp Vulkan snapshot read, 0.61 s to 0.03 s

    Odysseus · open source

    • The hub a self-hosted app for chat, agents, deep research, documents, email, calendar, notes and tasks
    • Agent loop native tool calling, plan mode and a tool-approval gate
    • MCP client connects to the servers Argus provides
    • Personas and skills where Jarvis's identity and jobs live
    • Model endpoints local engines and API providers, swapped at runtime
    • Cookbook downloads and serves local models

    The hub above is my own redrawing with sample conversations. This page runs no Odysseus code.

    Credits

    open-source projects

    Built with

    ProjectRole in ArgusLicence
    OdysseusThe hub Argus is built onAGPL-3.0
    Hermes Agent, Nous ResearchAgent, and the shared local model serverMIT
    llama.cppRuns the local modelMIT
    Qwen3.6-35B-A3BJarvis's local brainsee model card
    backtalkThe voice line my local brain plugs intoAGPL-3.0
    Kokoro-82MThe Daniel voiceApache-2.0
    faster-whisperSpeech to textMIT
    REAReverse-engineering toolsMIT
    ScraplingWeb-scraping toolsBSD-3-Clause
    VoiceStudioVoice toolssee project
    PaperclipPlugged into the hub; can manage HermesMIT
    Inspect AIRuns the evaluationsMIT
    Arize PhoenixTrace viewerElastic-2.0
    Talk to the public Jarvis← All my work