Research

Applied research for practical AI systems.

Readable papers on local agents, desktop control, and structured browser work.

Local LLM Routing

ZeroTurn

ZeroTurn is a deterministic pre-model routing layer for local agents. Its premise is narrow: when a user request is explicit, parameter-complete, and mapped to an allowlisted local action, the bridge should execute that action directly rather than spend a model turn deciding whether a known capability exists. This paper describes the tool's architecture, summarizes a local benchmark proof run, and discusses the implications of moving simple browser, file, and capability-inspection requests off the model path. In the committed benchmark, ZeroTurn deterministic dispatch reported a median latency of 29.641 ms, while a Codex model-turn classification baseline reported 3,789.852 ms, corresponding to a 127.86x median speedup for the benchmarked fast-path cases. The result supports a limited but useful claim: deterministic fast paths can materially reduce latency for explicit local actions, while ambiguous planning and reasoning should remain on the model path.

May 2026

Explore ZeroTurn

Desktop Control

Codesk Control

Desktop-control agents frequently interact with graphical operating systems through a loop of screenshot capture, visual interpretation, coordinate selection, and delayed verification. This loop is general, but it is often slow, brittle, and weakly structured. Codesk Control is a macOS control system that reverses the default order of interaction: it first extracts text and accessibility state, then acts through named keyboard shortcuts and accessibility actions, and only falls back to screenshots when the interface cannot be described textually. The system consists of a native Swift command-line tool and a native Model Context Protocol (MCP) server that exposes macOS state and actions as structured tools. In a live Codex desktop environment, moving the MCP path from a Node wrapper that spawned the CLI per request to a persistent native Swift MCP process reduced median state-call latency from 34.69 ms to 1.37 ms, a 25.3x speedup. Compared with representative AppleScript and screenshot probes in the same environment, the native MCP state tool was approximately 91x faster than an AppleScript front-application query, 105x faster than an AppleScript window-title query, and 56x faster than screenshot capture. We describe the design, implementation, benchmark methodology, use cases, security considerations, and broader implications of text-first desktop control for agentic systems.

2026

Explore Codesk Control

DOM Bridge

DOM-Native Browser Control

Language-model coding agents increasingly need to inspect, test, and operate web applications. The dominant interaction pattern still asks an agent to reason through screenshots, accessibility snapshots, browser automation wrappers, or repeated tool calls that mimic a remote human clicking through a page. This paper presents the Codex Web DOM Bridge, a lightweight local browser-control layer that exposes a compact command surface over the live Document Object Model (DOM): observe, act, wait, extract, search, and run. The bridge connects a local HTTP server, a browser content script, a command-line interface, and a Model Context Protocol (MCP) wrapper so that an agent can identify controls through labels, roles, selectors, and temporary handles, then execute multi-step DOM workflows in one browser hop. In a small demonstration benchmark, the bridge completed search-and-extract workflows with median latencies of 214–264 ms, compared with 1798 ms for a conventional in-app browser locator loop, a median speedup of 6.8–8.4x. The benchmark is intentionally modest, but it illustrates a broader design point: when a web page already exposes structured semantics, agentic control should use those semantics directly before falling back to pixels. The bridge has practical uses in local web testing, form filling, browser-based research, application QA, and AI-assisted development. It also sharpens important implications for consent, security, privacy, accessibility, benchmark design, and the future division of labor between visual automation and structured web control.

2026

Explore DOM-Native Browser Control

Image Bridge

Image Bridge is a local AI image-generation bridge for project work. It connects an already authenticated ChatGPT Images session in a dedicated Chrome profile to a command-line and model-callable interface, then writes generated PNG files to explicit project paths. The design is intentionally narrow: it is a human-supervised asset-production tool for local builds, not a hosted image-generation service or an API substitute. This paper describes the bridge pattern, its file-oriented workflow, and the operational boundary between exploratory asset work and production image infrastructure.

July 2026

Explore Image Bridge

Media Advertising

Media Advertising is a local AI-assisted campaign-production and publishing system for narrated image-to-video work. It constructs campaign assets, validates destination readiness, and keeps platform publication state separate through receipts. The package is designed around a practical rule: a queued job, a process exit, or an ambiguous browser result is not evidence that content was published. This paper examines the workflow as an engineering system for controlled multi-platform publishing rather than as a generic social-media automation promise.

July 2026

Explore Media Advertising

Video Creator

Video Creator is a local AI-assisted assembly tool for turning a written script, a shot list, image assets, optional music, and generated or attached narration into a vertical video. Its role is not to replace editorial judgment; it provides a reproducible media build path in which assets, timing, narration, and final encoding are recorded in a project manifest. This paper describes the pipeline, the role of local and optional speech providers, and the value of treating short-form video as a reviewable build artifact.

July 2026

Explore Video Creator

iPhone Control

iPhone Control is a state-aware local control layer for Apple iPhone Mirroring on macOS. It combines ordinary computer-use interaction with a small Swift controller for gestures that are unreliable through generic user-interface automation, including press-hold drags, waypoint routing, and folder drops. The controller discovers the visible mirrored-phone window at runtime, expresses gestures in that window's coordinate system, and stops when the observed state invalidates the action plan. This paper examines the design as a narrow reliability and safety layer rather than an autonomous device operator.

July 2026

Explore iPhone Control

Phonebot

Phonebot is a local voice edge for practical Mac Phone workflows. It is designed to answer unknown calls through the macOS Phone application, route audio through a local speech stack, apply an owner-phrase boundary, and expose a spoken control surface that can select a hosted or local language-model route. The packaged system deliberately excludes the broader legacy telephony and campaign stack. This paper describes the smaller workstation boundary, the local speech and audio requirements, and the implications of giving an owner a voice-controlled bridge to computer work.

July 2026

Explore Phonebot

Radio

Radio is a local-first system for inventorying nearby signals that the host may lawfully and passively observe. It combines desktop receiver workflows, optional local hardware adapters, user-controlled mobile ingest, salted local profiles, and evidence packages for authorized site surveys and personal environments. The product does not claim that ambient radio observations identify a person, and it excludes active interference, credential collection, forced pairing, spoofing, packet injection, and jamming. This paper explains the evidence model and why a consent-bounded observation system must preserve uncertainty.

July 2026

Explore Radio

Vision

Vision is a local-first visual-review system for two bounded workflows: reviewing video for timestamped vehicle and plate-related events, and recognizing enrolled trusted users from local still or webcam material. It keeps uploaded media, model weights, embeddings, settings, recordings, and retention controls close to the reviewing machine. The system provides analysis and review aids, not an autonomous safety decision or a blanket identification claim. This paper describes the local data model, event-recording boundaries, and the role of human review in practical visual workflows.

July 2026

Explore Vision

Mirror

Mirror is a local Wi-Fi presentation application for iPhone and iPad that sends a user-selected screen to a compatible Cast receiver. Its design uses a bounded local delivery path: screen media remains on the local network, the app does not operate a relay service for presentation content, and the current product does not stream audio. This paper describes the architecture as a local presentation workflow, the difference between local user-content transport and hosted software components, and the practical limits of the release surface.

July 2026

Explore Mirror

Radio Mobile

Radio Mobile is a user-started mobile companion for local observation and reporting. Its current scope is deliberately constrained to platform-appropriate Bluetooth, local-network and Bonjour discovery, foreground device context, on-device reports, local history, and optional user-initiated upload to a paired Radio desktop system. The application does not require an account and does not implement active interference, forced pairing, spoofing, credential collection, or packet injection. This paper explains why a mobile observation tool must make platform limits and user initiation central to its design.

July 2026

Explore Radio Mobile

AxleVision

AxleVision is a mobile visual-review application family for selected local camera and vehicle-related workflows. The documented mobile architecture supports camera setup, local stream handling, preview, adaptive inference, overlay rendering, event logging, annotated recording, retention controls, and a public-versus-managed configuration boundary. Production inference assets, broader codec support, physical-device testing, and vehicle-display authorization remain explicit release gates. This paper describes the design as a review tool with privacy and safety constraints, not as a substitute for mirrors, direct observation, or vehicle safety systems.

July 2026

Explore AxleVision