Skip to content

Channels

Jump to bottom
github-actions[bot] edited this page Apr 19, 2026 · 2 revisions

Channels — generated apps that push prompts back

Research preview. Available in the MCP server only, requires Claude Code CLI v2.1.80+ with a development flag. API is likely to change.

Neo's MCP server supports Claude Code Channels — a protocol that lets a running preview app push events back into the Claude Code session without waiting for user input. Claude automatically starts a new turn in response.

Two event types

  • runtime_error (system, always active) — Unhandled exceptions in the generated app are automatically pushed so Claude can react (usually: fix the bug and hot-reload).
  • user_trigger (user-defined via Ai.Trigger(prompt)) — The generated code decides when to push. The payload is a complete natural-language instruction; Claude executes it as if the user had typed it.

The Ai.Trigger API

The Neo.App.Api assembly is auto-referenced by every generated app. From inside a generated handler:

using Neo.App;

private void OnResearchClick(object sender, RoutedEventArgs e)
{
    var country = CountryCombo.SelectedItem?.ToString() ?? "(unknown)";
    Ai.Trigger(
        $"Research vacation destination {country}: best time to visit, " +
        $"top 3 sights, typical prices. Write the result with set_property " +
        $"into the TextBlock named 'resultText'.");
}

Click the button → Ai.Trigger(...) sends the prompt via MCP channel → Claude starts a new turn → executes the research using its tools → writes the result back into the running app via set_property. No user input between the click and the answer.

Ai.ScheduleTrigger(TimeSpan delay, string prompt) does the same thing after a delay — handy for timer-based behaviours.

Example end-to-end

You say (in Claude Code):

"Build an app with a combobox of vacation destinations and a 'Research' button. When I click the button, research the selected country and write the result into a TextBlock named 'resultText'."

Claude calls compile_and_preview with code that includes the Ai.Trigger call above. The window appears. You click the button. Claude gets a user_trigger event, executes the research, calls set_property on resultText, and the text appears in the same window — still alive, still responsive.

What this enables

  • Self-healing apps. Runtime errors turn into fix-it prompts without you noticing.
  • AI as runtime. The generated UI is the thin shell; the behaviour is Claude.
  • Business logic as prompts. "If the user selects X, fetch Y from the web and render Z." No pre-coding required.
  • Scheduled behaviours. Chains of triggers can model polling loops.

Costs and risks

  • Every trigger costs an API call. Don't wire it to MouseMoved, TextChanged-on-every-keystroke, or anything high-frequency. User-initiated actions only.
  • Latency. Trigger → new turn → tool calls → UI update is seconds, not milliseconds. Fine for "do the research" — terrible for "bounce a ball".
  • Prompt injection. Any text the running app passes into a trigger ends up in Claude's context. Don't run untrusted generated code with channels enabled.
  • Single-window for now. Multi-window routing of channel events is a future exercise.
  • Claude Code only (for now). Claude Desktop / Cowork don't support channels yet.

Enabling it

Channels aren't on the approved allowlist yet. Start Claude Code with:

claude --dangerously-load-development-channels server:neo-preview

The server:neo-preview part matches the MCP server's name in your Claude config. If you registered the MCP server under a different name, change it accordingly.

Once running, describe an app and the event behaviour you want ("when I click X, do Y"). Claude will include Ai.Trigger(...) calls automatically.

Protocol references

  • Claude Code channels reference — protocol details and limitations
  • MCP Server — the surface that implements the MCP-side of the channel
  • Neo.App.Api source in this repo — the Ai.Trigger and Ai.ScheduleTrigger implementations

Why this isn't in the desktop host or the Web App: both rely on Named Pipes or WASM runtime respectively, neither of which has a natural "push prompt back to Claude" path. The MCP variant is the only one that sits inside Claude Code's process boundary via JSON-RPC. Moving channels to the Web App would require a persistent websocket from the backend into the Claude Code CLI, which is not trivial and not currently on the roadmap.

Clone this wiki locally