LOPI is a self-hosted AI helper that runs entirely on your own hardware — no cloud AI service, no third-party agent framework. It owns its inference runtime, its memory, and its tools, and it can carry out approved, sandboxed tasks under strict, server-enforced limits.
LOPI blends a friendly chat experience with a carefully bounded agent layer. It can search the live web, remember what matters, read your approved documents, and act — but only within limits the server owns and you approve.
Streaming conversation on local models.
Citation-aware answers, current-date awareness, echo repair, and a one-click "Continue" for length-stopped replies — all on local inference with no cloud APIs.
Grounded answers from live sources and local documents.
Controller-owned search prioritises live sources for changing questions and approved local documents for stable ones. Provider keys never touch the browser.
Local memory with review-gated document storage.
Local SQLite/FTS5 memory plus a separate document store with immutable versions, conflict handling, and rollback.
Bounded agent tasks with explicit opt-in.
One validated action at a time, a step limit, a repetition guard, and durable, replayable task events — all within server-enforced limits.
Every mutation confirmed and recorded.
Mutations pause for exact-action approval. Every task is recorded with redacted, ordered events. Restart never replays an unproven change.
LOPI runs several local GGUF models behind node-aware aliases and independent chat, decision, and final roles. Every model is revision-pinned, license-tagged, and paired with its own instruction profile.
| Model | Size | Role / use | Status |
|---|---|---|---|
| Qwen3 1.7B | 1.3 GB | chat, decision, final | Default — tuned 6/8 threads |
| Gemma 3 4B | 2.5 GB | established CPU middle | Own sampling profile |
| Qwen3 8B | 5.0 GB | slower, higher quality | Temperature capped via eval |
LOPI's roadmap expands capability only after the safety substrate beneath it is proven. Foundations are complete; the durable task runtime is live; sandboxed workflows come next.
Strict action protocol, bounded orchestrator, canonical tool registry, multi-model routing, agent streaming, control plane, and local knowledge store.
Owner-bound ordered task events, a detached runner with a desktop queue, replay/polling, routed cancellation, and restart-safe semantics.
Bounded output schemas and risk/idempotency metadata on the canonical registry — classifying existing tools before expanding the surface.
A per-task artifact sandbox and a research-and-report pipeline: bounded reads, one confirmed write, role-based synthesis, deterministic verification.