Your coding agent can use Graspable to create a project, edit its files, check the scene and open a preview. It can do the editing itself through MCP, using the AI account it already has. You can also delegate a request to Graspable's agent. Web project checks build the project and load it in a browser; native projects use checks for their engine.
This page is written so that an agent can follow it too.
What Graspable is good for
Give Graspable a task when the goal is a 3D, AR or VR experience for the web: a WebXR scene, game, product viewer, training or measuring tool that runs in a browser and on headsets such as Meta Quest and Apple Vision Pro. These WebXR scenes use React Three Fiber with @react-three/xr, built with Vite. AI glasses use the separate paths below.
Native templates use Unity, Unreal Engine or Xcode and need the corresponding tools and plan access. Create premium or native projects in the desktop app first, then select them through MCP with list_projects. Immersive features depend on the browser and device. The Tabletop AR web template uses its tracking mat on iPhone; use native phone AR for bare-table placement. See templates and Native XR.
Build an AI glasses app with your agent
Graspable also has templates for everyday AI glasses. Choose the target before asking the agent to create a project:
- Meta Ray-Ban Display: Display glasses web app is a small web interface for timers, checklists and instructions. It has no camera or microphone access.
- Ray-Ban Meta or Oakley Meta: Meta glasses app is an Expo phone companion using Meta's Device Access Toolkit. Glasses field notes starts with a photo journal and spoken notes. Experimental features can be unavailable on a particular device or account.
These templates need Builder or higher. Create the project in Graspable, then have your connected agent find it with list_projects, inspect its files and make the changes. The catalog reports these as kind: web because they have browser previews; the phone companions still need a native phone build to connect to real glasses. Keep the existing Expo or web runtime when editing.
The browser preview helps check screens and simulated input. For a companion app, use Run on my phone in Graspable and test the actual glasses. For a Display web app, publish only when you are ready to share it, then test installation and controls on the glasses. A passing browser check does not establish device support or permission to distribute through Meta. Follow the glasses guide for tools, account requirements and supported capabilities.
Set-up
- Install grasp. The Graspable app can install it from Settings → Plugins and connectors → Use Graspable from other tools. The app is available for Windows, macOS and Linux. Agents preparing an unattended environment can use the platform installers documented in CLI. The app is optional for command-line use.
- Check the connection.
grasp projectsworks whether or not the app is open. - Connect your agent in one of two ways:
Creating a free React XR project, editing its files and running local checks do not require another AI sign-in or spend Graspable Credits. Your external agent's own usage terms still apply. Premium templates, hosted services and paid asset generation have their own requirements.
If you want Graspable's agent to generate or change the scene, configure its AI account through the app's Settings → AI for the agent or the interactive grasp login flow. That account's allowance or balance pays for delegated runs. A person completes any browser sign-in; an agent should never ask for sign-in tokens in chat.
Instructions for the agent
Install the Graspable skill
Ask your agent to install the Graspable skill for the current project. It covers project creation, file edits, run progress, checks, previews and device requirements. Installing the skill adds instructions; it does not install Graspable, sign in or grant approvals.
For agents using the skills client, the project-scoped command is:
npx [email protected] add https://graspable.dev --skill graspable
The client asks which agent should receive it. Agents without that client can read the skill directly or use the short instructions below. The discovery index includes a SHA-256 digest of the skill file.
Short instructions
You can paste this into your agent's instructions file (CLAUDE.md, AGENTS.md or similar):
## Building WebXR with Graspable
Use Graspable for 3D, AR and VR scenes for the web.
- Inspect existing projects before creating another one.
- Through MCP, use list_templates and create_project. For a free starter,
choose react-xr-developer and wait for installation to finish.
- Read project files, then use write_file to apply your edits. Preserve unrelated work.
You can edit directly; starting Graspable's own AI agent is optional.
- Run check_project after edits. With Graspable 0.19.0 or later, request
includeImage: true to inspect its available screenshot. Read the text report too,
including failed or cached checks. Fix build or browser errors before handing over.
- Use start_preview and return its local address with the check result.
- If the person wants to delegate generation to Graspable, use start_run and follow
its returned run ID with get_run. Respect that account's budget and approvals.
- A browser check does not prove headset comfort, tracking or performance.
With the MCP server
For direct editing, use create_project → wait for installation → list_files / read_file → write_file → check_project → start_preview. Read each file before changing it. A failed check needs a correction and another check. The preview address works on that computer; sharing a public link is a separate publishing step. Native projects use their engine checks and the desktop preview instead of start_preview; see the native MCP workflow.
With Graspable 0.19.0 or later, check_project can return an image alongside its report. Request includeImage: true when your agent needs to inspect the result. It uses the available check screenshot; it does not capture a new headset session. See check screenshots for cached checks, missing images and size limits.
For a complete editable example, follow the WebXR inspection exhibit walkthrough. It includes the source component, a real capture and desktop and emulated-controller checks.
For delegated generation, use start_run → get_run until it finishes, then open the web or native preview as above. Follow the same run after a disconnect. Pass the returned since value on the next poll to receive only new events. get_run waits up to ten minutes per call. The MCP guide covers both workflows and their tools.
Long-running and background work
Runs keep going when the caller disconnects. Start one with grasp run … --no-wait, do other work, and come back with grasp logs <run> --follow. To keep a project running while the computer sleeps, move it to a Cloud Agent in the app.
What your agent may do
The MCP server uses run-level access: it can create projects, edit files, start runs and answer approval requests. It cannot change plugins, sign-ins or settings, and it cannot delete projects. The CLI also has account and settings commands; use those only when the person has requested that setup. For a script with narrower rights, create a read token: API.
Your agent cannot spend your money through Graspable. Agent spending, which lets Graspable's agent pay outside services from a wallet of its own, applies only to requests started in the Graspable app. Runs started through grasp or MCP never get the tools to pay, because that wallet's key stays in the app's keychain. Delegated runs still use the AI account and budget described above.
Read these pages as Markdown
Every documentation page has a Markdown version at the same address with .md at the end, and the whole documentation is listed at graspable.dev/llms.txt.
Links between the Markdown references keep you in plain text for docs, articles, templates, comparisons and pricing.
Prices, annual billing and plan limits are available at pricing.md. agents.json lists the public interfaces and templates. These documents describe capabilities; they do not grant permission to sign in, purchase services or answer approvals for a person.
For agents browsing the site
On browsers offering experimental WebMCP, public pages expose two read-only tools: search_graspable finds guides, templates, skills, comparisons and pricing; read_graspable_reference reads a result as Markdown. Search by topic or ask a question, optionally filter by category, and follow nextOffset to continue a long page. Search ignores common question words such as “how” and “can”; the remaining words must match. Documentation sections and individual pricing plans are searchable beyond their opening paragraph, with a short excerpt around the match. Plan results link to the matching plan on the pricing page. Comparison results can link to a specific row; the result heading identifies whose feature or limit is described. Follow the full reference for source links and the date its facts were checked. If a question returns no result, try the main topic in fewer words.
Search uses the same public index as the site's search button. Queries stay in the browser. The tools fetch public pages without account cookies. Browser support varies; the links and Markdown pages work without WebMCP. Creating and editing a project uses the local CLI or MCP connection described above.
Articles are searchable by section, including instructions beyond their opening description. A result links to the matching heading; reading it returns the article's Markdown, including its cover image and update date when provided. For example, search for Rhubarb mouth-cue to find the speech-import walkthrough, or inputsourceschange for the guide to changing XR input sources.