Local LLMs & offline use

ContextIQ is model-agnostic: a context layer, not a model. It parses your repo, ranks what matters, and emits a plain-text / Markdown (or JSON) pack. It never calls an LLM, so it works with any model — cloud or local.

Pipe a pack into a local model

tokengraph context "explain the auth flow" | ollama run llama3

No API key, no telemetry, nothing leaves the box. The same works with any runner that reads a prompt on stdin — llama.cpp, LM Studio, vLLM.

Why this matters

• Privacy — your code never goes to a third party just to build context.
• Cost — a tight pack keeps even large local-model prompts fast and cheap.
• Portability — swap models freely; the context layer doesn't change.

Pair it with the trust gates

The validate / judge / verify gates are also fully local — so you can run the entire retrieve → answer → check loop offline, with a local model doing the answering.

Model-aware costing (when you do use a cloud model)

tokengraph cost              # estimate input+output USD before sending
tokengraph cost --compare    # pick the cheapest model that's still sufficient

Covers GPT / Claude / Gemini / Llama pricing — deterministic and local.

Next steps

• Wire it into an editor: MCP server
• Keep prompts tight: prompt-score, dedupe, summarize-chat in the CLI reference