occ / node

OCC Node

The local runtime at the heart of OCC. Download it, run it, and join the network. The node handles model selection, private retrieval, peer collaboration, and broker communication — all from your own hardware.

how it works

01

Chat or retrieval — instant classification

Before answering, a lightweight classifier decides whether the query is a general conversation (CHAT) or requires knowledge retrieval (DELIBERATE). Chat queries are answered immediately by the local model. Retrieval triggers the full pipeline below.

02

Centralized, tamper-proof knowledge

All official knowledge packs live on the OCC server — curated, reviewed, and approved by the community. Because knowledge no longer resides on user machines, it cannot be altered or poisoned by malicious nodes. Trust is built into the architecture.

03

Private retrieval — your query never leaves your PC

When retrieval is needed, the node downloads the indexes of all server packs and runs the semantic search locally. Only a list of the specific wiki pages it needs is sent back to the server. The query itself is never transmitted — full privacy by design.

04

Expert Draft + Peer Critic

The server returns the requested wiki pages along with a list of available peers. The local LLM writes a first Expert Draft. If a remote peer with superior hardware is online, it receives the draft, acts as Critic — reviewing it and flagging gaps — and sends back the critique. The local model then synthesizes the final response. If no stronger peer is available, the Critic step also runs locally.

05

OpenRouter support

Want faster inference? Route your own queries through OpenRouter with your personal API key — currently supporting qwen3.5-9B and qwen3.5-35b-a3b, at minimal cost. Crucially, your GPU still serves the network as a Peer Critic for remote users regardless. OpenRouter only affects your own queries — you never pay for others.

06

Forge & private packs

Use Forge to build knowledge packs from your own documents, personal data, or proprietary information. Private packs stay on your machine and are never uploaded. Run /local on to use them exclusively. When ready, you can submit a pack for community review and publication to the official server.

model tiers

OCC detects available VRAM at startup and selects the appropriate model automatically. All tiers use the Qwen family with adaptive context optimised for your hardware, and support multimodal input.

TierVRAMModelQuantSize
MicroCPUqwen3.5:2bQ4_K_M
Small4 GBqwen3.5:4bQ4_K_M3.4 GB
Mid8 GBqwen3.5:9bQ4_K_M5.97 GB
Large16 GBqwen3.5:9bQ8_09.53 GB
XL24 GBqwen3.5:9bBF1619 GB
Server S32 GBqwen3.6:27bQ4_K_M16.8 GB
Server L80 GBqwen3.6:27bBF1656 GB

OCC's promise is to deliver frontier-grade answers using small, affordable models — by combining reliable pack knowledge with distributed peer review rather than relying on raw model size.

built-in tools

Node exposes a set of tools available in all LLM calls. File and code operations are strictly sandboxed to /workspace — no access to the rest of the filesystem.

web_search

Search the web via DuckDuckGo. Must be requested explicitly — this is an intentional guardrail that forces the model to rely on curated pack knowledge by default rather than fetching unverified results on every query.

fetch_url

Fetch and read the content of a URL.

read_file

Read a file from the sandboxed /workspace directory.

write_file

Write or create a file inside /workspace.

list_files

List files and directories inside /workspace.

run_code

Execute code with a 30-second timeout, sandboxed to /workspace. No access to the rest of the filesystem.

commands

/clear

Clear conversation history and reset context

! <query>

Force network mode — consult peers regardless of local knowledge

/packs

List all loaded packs and domains

/peers

Show active peer nodes on the broker

/status

Show current config

/local on

Use local packs only — for private Forge packs

/local off

Use server packs (default)

/unload

Unload model from VRAM

/load

Reload model into VRAM

/openrouter on

Switch to OpenRouter (if configured)

/openrouter off

Switch to local model

What's next

The browser GUI is the current interface, served locally at localhost:7891. A native desktop app is planned for a future phase. The underlying node runtime will remain unchanged.