/**
* ContextStream Media-Aware Hook
*
* Detects media-related prompts (video, clips, Remotion, image, audio) and
* injects context about the media tool.
*
* Usage:
* npx @contextstream/mcp-server hook media-aware
*
* Input (stdin): JSON with prompt or session messages
* Output (stdout): JSON with hookSpecificOutput containing media tool guidance
* Exit: Always 0
*/
const ENABLED = process.env.CONTEXTSTREAM_MEDIA_HOOK_ENABLED !== "false";
// Media patterns (case-insensitive)
const PATTERNS = [
/\b(video|videos|clip|clips|footage|keyframe)s?\b/i,
/\b(remotion|timeline|video\s*edit)\b/i,
/\b(image|images|photo|photos|picture|thumbnail)s?\b/i,
/\b(audio|podcast|transcript|transcription|voice)\b/i,
/\b(media|asset|assets|creative|b-roll)\b/i,
/\b(find|search|show).*(clip|video|image|audio|footage|media)\b/i,
];
const MEDIA_CONTEXT = `[MEDIA TOOLS AVAILABLE]
Your workspace may have indexed media. Use ContextStream media tools:
- **Search**: \`mcp__contextstream__media(action="search", query="description")\`
- **Get clip**: \`mcp__contextstream__media(action="get_clip", content_id="...", start="1:34", end="2:15", output_format="remotion|ffmpeg|raw")\`
- **List assets**: \`mcp__contextstream__media(action="list")\`
- **Index**: \`mcp__contextstream__media(action="index", file_path="...", content_type="video|audio|image|document")\`
For Remotion: use \`output_format="remotion"\` to get frame-based props.
[END MEDIA TOOLS]`;
interface HookInput {
prompt?: string;
session?: {
messages?: Array<{ role: string; content: string | Array<{ type: string; text?: string }> }>;
};
hookName?: string;
}
function matchesMediaPattern(text: string): boolean {
return PATTERNS.some((pattern) => pattern.test(text));
}
function extractPrompt(input: HookInput): string {
// Direct prompt field
if (input.prompt) {
return input.prompt;
}
// Extract from session messages (last user message)
if (input.session?.messages) {
for (let i = input.session.messages.length - 1; i >= 0; i--) {
const msg = input.session.messages[i];
if (msg.role === "user") {
if (typeof msg.content === "string") {
return msg.content;
}
if (Array.isArray(msg.content)) {
for (const block of msg.content) {
if (block.type === "text" && block.text) {
return block.text;
}
}
}
break;
}
}
}
return "";
}
function detectEditorFormat(input: HookInput): "claude" | "cline" {
if (input.hookName !== undefined) {
return "cline";
}
return "claude";
}
export async function runMediaAwareHook(): Promise<void> {
if (!ENABLED) {
process.exit(0);
}
// Read stdin
let inputData = "";
for await (const chunk of process.stdin) {
inputData += chunk;
}
if (!inputData.trim()) {
process.exit(0);
}
let input: HookInput;
try {
input = JSON.parse(inputData);
} catch {
process.exit(0);
}
const prompt = extractPrompt(input);
// Only inject media context if prompt matches media patterns
if (!prompt || !matchesMediaPattern(prompt)) {
process.exit(0);
}
const editorFormat = detectEditorFormat(input);
if (editorFormat === "claude") {
// Claude Code format
console.log(
JSON.stringify({
hookSpecificOutput: {
hookEventName: "UserPromptSubmit",
additionalContext: MEDIA_CONTEXT,
},
})
);
} else {
// Cline/Roo/Kilo format
console.log(
JSON.stringify({
cancel: false,
contextModification: MEDIA_CONTEXT,
})
);
}
process.exit(0);
}
// Auto-run if executed directly
const isDirectRun = process.argv[1]?.includes("media-aware") || process.argv[2] === "media-aware";
if (isDirectRun) {
runMediaAwareHook().catch(() => process.exit(0));
}
