Authoring generative UI over AG-UI

CAO exposes an AG-UI stream (GET /agui/v1/stream) that any dashboard — CopilotKit, the AG-UI Dojo, or a plain EventSource — renders without CAO-specific code. As an agent you can push a declarative UI intent onto that stream with the emit_ui MCP tool. The operator sees a rendered card, not raw text — and because every provider's intents render uniformly, they can't tell (and don't need to) which CLI agent produced which card.

The surface must be enabled on the server (CAO_AGUI_ENABLED=true or CAO_MCP_APPS_ENABLED=true — the two surfaces share one event source). When it is disabled, emit_ui returns {"ok": false, "reason": "AG-UI surface disabled…"} — treat that as a no-op, not an error.

Safety model (why this is always safe to call)

You may emit only a closed allow-list of named components with JSON props. There is no HTML, no script, no eval, no iframe. The intent is validated server-side against the allow-list before it reaches the stream:

• An off-list component (e.g. iframe, script) is refused — the tool raises a ValueError; nothing is rendered.
• props must be JSON-serializable and are bounded to 8 KB — an oversized or non-serializable payload is rejected at the emit_ui boundary (HTTP 400, the tool raises a ValueError), so a bad payload never reaches the bus.
• If the AG-UI surface is disabled on the server, the tool degrades gracefully (no error) — so calling it is never fatal.
• The AG-UI stream is metadata-only by contract: never put message bodies, credentials, or file contents in props. Reference paths, not contents.

The tool

emit_ui(component: str, props: dict) -> {"ok", "event_id", "component"}

component must be one of: approval_card, choice_prompt, diff_summary, progress, metric, agent_card.

When to use which component

Props below are what a conformant client renderer will display; unknown extra keys are ignored, not refused.

Examples

# Gate a risky action on human approval.
emit_ui("approval_card", {
    "title": "Deploy to production?",
    "detail": "3 files changed, 1 DB migration",
    "risk": "high",
})

# Ask the operator to choose.
emit_ui("choice_prompt", {
    "question": "Which base branch?",
    "choices": [{"label": "main", "value": "main"},
                {"label": "release", "value": "release"}],
})

# Summarize a change set.
emit_ui("diff_summary", {
    "title": "Refactor auth",
    "files": [{"path": "security/auth.py", "additions": 74, "deletions": 3}],
})

# Show progress / a metric / your identity.
emit_ui("progress", {"label": "Indexing repository", "value": 0.42})
emit_ui("metric", {"label": "tokens used", "value": 12840, "unit": "tok"})
emit_ui("agent_card", {"name": "reviewer", "provider": "claude_code", "status": "working"})

L2 constructs (Phase 2)

The AG-UI surface also exposes L2 constructs — higher-level projections that fold the raw event stream into structured views. As an agent you don't author L2 constructs, but you should know they exist because your emit_ui intents feed them:

• SupervisorDashboardStream — folds STATE_SNAPSHOT/STATE_DELTA + your agent_card emits into a live fleet hierarchy view.
• MultiAgentSessionTimeline — reconstructs delegation/message timeline from TOOL_CALL lifecycle events.
• AgentHandoffWithApproval — the full interrupt lifecycle: provider prompt → reason classification → interrupt → approve/deny/edit → delivery.
• CrossProviderStateSync — convergence proof across providers.

The run plane (POST /agui/v1/run) streams these as stock AG-UI wire frames. Interrupts (approval prompts) route through POST /agui/v1/interrupts/{id}/resume.

For details: references/l2-constructs.md and references/run-plane.md.

Gotchas

1. Emitting to a disabled surface — if CAO_AGUI_ENABLED is unset, emit_ui returns {"ok": false} gracefully. Don't treat this as an error or retry — it's a no-op by design. The fix: always check ok in the return but never fail on it.
2. Props over 8 KB are rejected — the tool raises a ValueError and nothing renders. The fix: reference file paths instead of embedding content. Keep props to metadata (paths, counts, labels).
3. No HTML sink exists — strings in props render as plain text. Attempting to smuggle markup through props (e.g. <script>, <iframe>) won't render and looks broken. The fix: use structured props, not markup.
4. One intent per meaningful moment — emitting a progress card on every token or tool call floods the stream and degrades client rendering. The fix: emit at milestones (start, 25%, 50%, 75%, done) or once per logical phase.
5. approval_card is display-only today — it gives the operator an approve/reject affordance in the dashboard, but the action routes to the dashboard's command surface, not back to you. The fix: pair it with your provider's own wait-for-input mechanism (e.g. Kiro's trust prompts, Claude Code's permission dialog).
6. Off-list components are refused server-side — the allow-list is fixed (approval_card, choice_prompt, diff_summary, progress, metric, agent_card). A typo or new component name returns HTTP 400. The fix: use only the six listed names; check spelling.

Verifying locally

# 1. Server with the surface on
CAO_AGUI_ENABLED=true uv run cao-server

# 2. Watch the stream (SSE frames print as they arrive)
curl -N 'http://localhost:9889/agui/v1/stream'

# 3. Emit from anywhere (the MCP tool does exactly this)
curl -sX POST http://localhost:9889/agui/v1/emit_ui \
  -H 'Content-Type: application/json' \
  -d '{"component":"progress","props":{"label":"demo","value":0.5}}'

A GENERATIVE_UI frame with your component appears on the stream; an off-list component is refused with HTTP 400.

See also

• examples/ag-ui/ag-ui-dashboard/ — a runnable demo (run.sh + showcase.sh) that drives all six components live and shows the off-list refusal.
• docs/agui.md — the AG-UI stream and generative-UI reference.
• cao-mcp-apps skill — operate and extend the MCP Apps surface that renders your emit_ui intents inside host dashboards (Claude Desktop, VS Code, etc.).
• mcp-apps-builder skill — build new MCP App views that consume the AG-UI stream your emits feed into.