A visual context layer

Visual context for
AI workflows.

Point, mark, and speak. Canvas Prompt carries your mark, your words, the object, and the moment—so your AI gets the context it needs. Works alongside AI terminals, CLI tools, and agents. Start with Codex.

INSTALL FOR CODEX ↓EXPLORE THE OPEN CONTEXT PACKAGE ↓
PUBLIC ALPHA · LOCAL CONTEXT · AI WORKFLOWS · START WITH CODEX
landing-page.draft00:08.4
PROJECT / HOME PAGE

Make AI work where your thinking happens.

A spatial surface for marks, references, spoken feedback, and the next decision.

“Move this up a little.”

this → “Make AI work…”

Recommended integration: Codex

Start with Codex.

Canvas Prompt saves context on your device. After a round, ask your current Codex task to read the latest package and continue from it. Other AI terminals, CLI tools, and agents need an adapter.

# Add the Canvas Prompt marketplace
codex plugin marketplace add https://github.com/dlxeva/canvas-prompt

# Install Canvas Prompt
codex plugin add canvas-prompt@canvas-prompt

Codex: complete public integration · Other AI hosts: open canvas + adapter + real install validation. Open a new Codex task after installing or updating to load the plugin.

Your first round

Start with one correction.

You do not need a finished diagram or a perfect prompt.

01 / PLACE

Put one image, page, or rough draft on the canvas.

02 / MARK

Circle the spot. Say or type what should change.

03 / CONTINUE

Finish the round, then type Continue with the canvas context in your current Codex task. Codex restates its reading before material changes.

PDF · PPTX · INTERACTION REVIEW

Review the actual work, page by page.

Choose Interactive Review inside Canvas Prompt. Mark a local PDF or PPTX, speak beside the page, and send the anchored evidence to your AI. The original stays read-only.

01 / OPEN

Choose a document.

Click Interactive Review in the top workspace switcher, or drop a PDF/PPTX onto the canvas.

02 / MARK

Keep the page anchor.

Turn pages, zoom, draw, circle, point, and speak. PDF pages render locally; PPTX uses a review-only PDF derivative when LibreOffice is available.

03 / CONTINUE

Let AI restate first.

Export the review package and ask Codex to read it. The source stays unchanged, and execution still requires your confirmation.

The immediate use

Feedback should not have to leave the work.

Visual feedback often becomes a screenshot, a long explanation, and a guess about which “this” you meant. The canvas can retain the reference itself.

01 / THE HANDOFF

Screenshot → chat

Mark a place. Move to a conversation. Reconstruct the location, object, and instruction in text.

“the title near the top…”
02 / THE CANVAS

Mark → speak → continue

Circle the object and say what you want to change. The reference stays tied to the surface while your AI gets the context it needs.

“move this up a little”
Two ways in

See the interaction.
Then the wider idea.

Start with the concrete moment: point, speak, and keep the reference. Then see why a full working path can matter too.

Point. Speak. Continue.

START HERE · 00:25

Your rough work is the prompt.

DEEPER WORK · 00:45
WATCH THE FULL LIVE DEMO · 03:50 ↗
The same canvas, more room

A correction is where most people start.
Some work needs the whole path.

When the question itself is still forming, the marks, hesitations, rejected routes, and spoken reasons are part of what the next AI needs.

A session replay

Watch a judgment take shape.

The replay reconstructs a 76-second session: alternatives are placed, one is rejected, and a final direction is marked on the canvas.

● REC · 00:00 → 01:16REPLAY / ONE SESSION
Which product?CodelessH2BiPadgo with this ✓costly, hard to build
WHAT HAPPENED: candidates · pause · spoken reason · rejection · callTHE FINAL CANVAS ALONE DOES NOT HOLD ALL OF IT
What the chat box loses

By the time you hit send, the compression has already happened.

  • cross-out · “iPad” rejected at 00:55revision
  • 6.2s pause before the rejectiontiming
  • two candidates kept, one branched asidestructure
  • spoken reason: cost & development difficultyspeech
  • final circle: “go with this ✓”decision
WHAT THE AI RECEIVED
Help me decide which product to go with.
↑ everything flattened into one sentence

The crossed-out paths, pauses, and spatial relationships
do not fit inside a normal prompt.

{
  "meta": { "format": "compact_timeline_aware" },
  "semantic_events": [
    { "type": "reject_path", "reason": "cost, dev difficulty" },
    { "type": "confirm_mainline" }
  ],
  "object_states": {
    "survived": ["Codeless", "H2B"],
    "rejected": ["iPad"]
  },
  "evidence_refs": ["[email protected]"]
// each claim stays linked to a session moment
}
The Prompt Package

The output is not a summary.
It’s evidence.

Canvas Prompt organizes the working material of a session into context another AI workflow can inspect and continue from.

semantic_eventsWhat you concluded—and when.
object_states.rejectedWhat you crossed out remains on record.
evidence_refsEvery claim links back to an observable moment.
Thinking sovereignty

You define the question.
You decide what stays.

Canvas Prompt keeps the working material of your judgment available to the AI that joins next. The question, the paths worth keeping, and the final call remain yours.

01You frame the question.
02You keep or reject the paths.
03You make the call.
Design partners

Bring one real question.

We are looking for deep AI collaborators: people who already work with AI and still reach for a sketch, a whiteboard, or a rough diagram when the work is not clear yet. Bring one live problem. One session. One export.

BECOME A DESIGN PARTNER ↗
Canvas Prompt

Put a complete thinking session into your AI workflow.

Start with one image and one correction. Stay when the work needs more room than a chat box can hold.

BECOME A DESIGN PARTNER ↗