OmniLAB / LocalAgentBridge

Local AI assistance for coding, with explicit supervision and review.

OmniLAB connects Codex to local models for focused explanations, code reviews, tests and bounded fixes. It keeps exact model choices, durable jobs and saved evidence together, so a supervisor can inspect and accept each contribution.

Get started · Browse the guides · Open an application · Documentation website

Development source: 0.5.0.dev0. Latest released version: v0.4.1. Merging source does not qualify experimental engines, upgrade an installed runtime or publish a release. See current status.

Start here

Start with a focused advisory question and selected source. Use an exact installed model, inspect its saved answer against the source, then record your independent decision. Finished generation is not accepted work.

Windows is the tested desktop target. The candidate needs Python 3.11+, a private LAB configuration/data store, Ollama and an installed completion model. The dock also needs Tcl/Tk; managed Codex reviews need an authenticated Codex CLI. Local trusted review is supported. Follow the quickstart for installation and commands.

Related MCP server: claude-token-saver-mcp

How it works

Three task lanes share explicit scope, evidence and independent acceptance:

flowchart TD
    C["Codex: plan and supervise"] --> L["OmniLAB: validate, route and retain evidence"]
    L -->|"return scope and reasons"| D["Codex direct"]
    L --> A["Local advisory"]
    L --> H["Local implementation: OpenHands SDK"]
    A <-->|"direct provider call"| O["Ollama"]
    H <-->|"restricted inference relay"| O
    H -.->|"separate qualification required"| S["Strata: experimental"]
    D --> R["Codex: independent review and acceptance"]
    A --> R
    H --> R

• Codex direct returns the scope and reasons to the supervisor. LAB does not make an automatic cloud call or silently change models.

• Local advisory asks a local model for bounded analysis. It works independently of OpenHands qualification.

• Local implementation gives OpenHands a restricted read/edit/test loop. The networkless sandbox reaches inference through a trusted LAB relay; the supervisor checks the candidate and protected acceptance results.

Ollama and Strata are alternative inference providers. Strata is disabled pending qualification. LAB uses the OpenHands SDK, which does not install an OpenHands browser application. The workflow guide explains the routing decisions; its execution sequence shows transport, permitted tools and verification.

Current status

These personal profiles are dated, narrow observations; they are not public hardware requirements or permission for arbitrary tasks. Other operating systems are not certified. Independent Janas experiments have their own research and access guide.

Documentation

The guides render directly on GitHub. The documentation website presents the same source with search and sidebar navigation.

See the documentation index for the retained workflow interfaces, benchmarks, decisions and contribution history.

Privacy and evidence

Task snapshots, local inference, SQLite records and candidate artifacts stay on the configured machine. Selected source, findings, diffs and verification evidence may reach Codex's configured cloud service during planning or review. Check that content and your provider settings for sensitive work. The selective sandbox has no external network or host-home mount; trusted LAB mediates inference.

Models propose work. They cannot grant permissions, approve their own results, merge, publish or install another engine. Local token counts do not establish cloud savings; that claim requires an equivalent accepted baseline with complete overhead and usage accounting. The comparison protocol separates synthetic instrumentation checks from live evidence.

Repository map

Contributors should read CONTRIBUTING.md and AGENTS.md. OmniLAB is MIT licensed; model and dependency licenses are separate.