PixelOkay
Local-first AI
Tools
ON-DEVICE • GUIDED AGENT WORKSPACE

Agentic AI

Plan tasks, inspect files, and ask a local AI to propose changes. Review each diff, apply only what you choose, undo safely, and export your workspace. AI runs on your device.

✳ Runs on your device◉ You approve sensitive actions◈ Review diffs & undo⬇ Export local ZIP
AI model not loadedImport and inspect files anytime. AI tasks need a locally loaded model.
OPTIONAL MODEL CACHE CLEANUP

Delete downloaded model files?

Recognized WebLLM model caches are selected for removal.

Other PixelOkay AI tools sharing these cached files may have to download them again. Imported files, saved projects and general website data are not targeted. Unknown caches may remain.
STEP 2 / YOUR TASK

What would you like to do?

Describe the desired outcome. Guided mode lets you approve or edit the plan before any action.

Need an idea?

Choose how your AI helps

Guided mode creates a plan you can edit before the agent acts. Tool actions are still restricted and sensitive actions require approval.

Drop UTF-8 text, data, HTML, or source-code files here, or choose files from your device.

Max 12 files, 2 MB each, 6 MB combined. No uploads to PixelOkay. Files are editable copies in this tab, not your originals.

    Save or restore your workspace

    Keep an editable, local project including your imported documents and partial work. Project files may contain private data.

    STEP 3 / RESULTS

    Agent workspace

    Waiting for model
    ✦

    First, load your AI model

    Click Set up AI model, then describe your task and choose a workflow. You can browse and edit imported files without AI.

    1 · Load AI2 · Describe task3 · Review changes
    STEP 4 / FILE CHANGES

    Review & edit your workspace

    The AI only stages suggestions. Your original files on disk are never modified. Review changes here and click Apply to update the in-tab copy.

    No changes yet
    Quick edit without AI — find and replace

    Stage a precise replacement in an imported text file. You can review the diff before applying it.

    No proposals yet. Import a code or text file, then ask the agent to suggest a change—or use Quick edit above.

    Applied edits are kept in this browser tab. Export the workspace ZIP or download an individual file to save changes.

    BUILT FOR PRACTICAL TASKS

    Tools under your control

    01

    Work with documents

    Search, read and compare imported reference files without giving the agent access to your device’s filesystem.

    02

    Calculate & analyze CSVs

    Use a validated calculator, grouped CSV summaries and a local JSON inspector to ground conclusions.

    03

    Read a public web page

    Approve each public HTTPS read. Sites may disallow cross-origin access, and the tool cannot bypass that.

    04

    Review & resume projects

    Inspect recorded tool evidence, edit notes, resume saved projects, and export clearly marked reports.

    How Agentic AI works

    1. Load the AI model onto your device. 2. Describe a task and optionally attach your files. 3. Choose Ask, Plan, Guided, or Auto mode; guided mode waits for your plan approval. 4. Review action requests and evidence. 5. Inspect proposed file diffs, apply only the changes you trust, and export your workspace or project.

    The model may make mistakes, especially smaller models. Saved projects contain imported file contents and should be treated as sensitive. This is a bounded task assistant, not an unrestricted autonomous computer operator. Browser security, site CORS rules, and available hardware limit its capabilities.

    Using Agentic AI

    Use this local AI workspace to analyze imported files, plan bounded tasks and prepare text or code changes for your review. Choose a supported runtime and load a model before starting. Proposed file changes remain staged until you approve them.

    Use Agentic AI in three steps

    1. Select a supported model/runtime and choose Download / load AI. Runtime or model files are fetched only after you start setup.
    2. Import up to 12 UTF-8 text, data or source-code files (2 MB each and 6 MB total), describe the outcome you want, then choose Ask, Plan, Guided or Auto mode.
    3. Review the plan, tool evidence and proposed diffs. Approve actions where requested, apply only the file changes you accept, then undo if needed or export the workspace/project.

    Example task

    With a sales CSV imported, try: Summarize monthly sales, calculate the total from units × price, note any assumptions, and write a short Markdown report. Do not modify the source CSV.

    Limits and data handling

    Requires a compatible browser/device and WebLLM/WebGPU or browser Prompt API support; local model quality and performance vary. Public HTTPS page reads require per-action approval, browser CORS access, and are limited to 128 KiB and 12 seconds; redirects are rejected. There is no shell, arbitrary JavaScript execution, general browser control or unattended task runner. Source-code files are read as text, not executed. File proposals require explicit approval before they are applied. Project/checkpoint exports can include full imported file contents and action evidence, so inspect them before sharing. Runtime behavior and real model output remain unverified.

    Frequently asked questions

    Does Agentic AI run imported code or apply every suggested change?

    No. It inspects code as text. Proposed file changes are staged for review and require your approval before application; you can inspect diffs and undo changes.

    Can it read any website automatically?

    No. Each public HTTPS page read requires approval and is subject to the destination site's CORS policy. Redirects are rejected, and the tool cannot bypass browser security.

    Related tools