diff --git a/apps/web/src/app/(profile)/[username]/collections/[slug]/skills.md/route.ts b/apps/web/src/app/(profile)/[username]/collections/[slug]/skills.md/route.ts
new file mode 100644
index 0000000..30f28c3
--- /dev/null
+++ b/apps/web/src/app/(profile)/[username]/collections/[slug]/skills.md/route.ts
@@ -0,0 +1,192 @@
+/**
+ * GET /:username/collections/:slug/skills.md
+ * Generate AI-powered skills documentation for a collection
+ * Clean URL endpoint for machine-readable skills documentation
+ */
+
+import { prisma } from '@tpmjs/db';
+import type { NextRequest } from 'next/server';
+
+import { fetchMultiplePackageSources } from '~/lib/ai/package-source-fetcher';
+import { generateSkillsMarkdown } from '~/lib/ai/skills-generator';
+
+export const runtime = 'nodejs';
+export const dynamic = 'force-dynamic';
+export const maxDuration = 120; // Extended for source fetching + AI generation
+
+// Cache for 1 week
+const CACHE_DURATION_SECONDS = 7 * 24 * 60 * 60; // 604800 seconds
+
+type RouteContext = {
+  params: Promise<{ username: string; slug: string }>;
+};
+
+/**
+ * GET /:username/collections/:slug/skills.md
+ * Generate skills.md markdown for a collection
+ */
+export async function GET(request: NextRequest, context: RouteContext) {
+  try {
+    const { username: rawUsername, slug } = await context.params;
+    const username = rawUsername.startsWith('@') ? rawUsername.slice(1) : rawUsername;
+
+    // Check for cachebust query param
+    const cachebust = request.nextUrl.searchParams.has('cachebust');
+
+    // Find the user first
+    const user = await prisma.user.findUnique({
+      where: { username },
+      select: { id: true, username: true },
+    });
+
+    if (!user || !user.username) {
+      return new Response('User not found', {
+        status: 404,
+        headers: { 'Content-Type': 'text/plain' },
+      });
+    }
+
+    // Find the collection by slug belonging to this user
+    const collection = await prisma.collection.findFirst({
+      where: {
+        slug,
+        userId: user.id,
+      },
+      include: {
+        tools: {
+          include: {
+            tool: {
+              select: {
+                id: true,
+                name: true,
+                description: true,
+                inputSchema: true,
+                package: {
+                  select: {
+                    npmPackageName: true,
+                    npmVersion: true,
+                  },
+                },
+              },
+            },
+          },
+          orderBy: { position: 'asc' },
+          take: 50, // Limit tools for performance
+        },
+      },
+    });
+
+    if (!collection) {
+      return new Response('Collection not found', {
+        status: 404,
+        headers: { 'Content-Type': 'text/plain' },
+      });
+    }
+
+    // Only return if public
+    if (!collection.isPublic) {
+      return new Response('This collection is not public', {
+        status: 403,
+        headers: { 'Content-Type': 'text/plain' },
+      });
+    }
+
+    // Check cache if not busting
+    if (!cachebust && collection.skillsMarkdown && collection.skillsGeneratedAt) {
+      // Check if cache is still valid (within 1 week)
+      const cacheAge = Date.now() - collection.skillsGeneratedAt.getTime();
+      if (cacheAge < CACHE_DURATION_SECONDS * 1000) {
+        return new Response(collection.skillsMarkdown, {
+          status: 200,
+          headers: {
+            'Content-Type': 'text/markdown; charset=utf-8',
+            'Cache-Control': `public, s-maxage=${CACHE_DURATION_SECONDS}, stale-while-revalidate=86400`,
+            'X-Cache': 'HIT',
+            'X-Cache-Age': Math.floor(cacheAge / 1000).toString(),
+          },
+        });
+      }
+    }
+
+    // Extract tools data
+    const tools = collection.tools.map((ct) => ({
+      id: ct.tool.id,
+      name: ct.tool.name,
+      description: ct.tool.description,
+      packageName: ct.tool.package.npmPackageName,
+      inputSchema: ct.tool.inputSchema,
+    }));
+
+    if (tools.length === 0) {
+      return new Response('Collection has no tools', {
+        status: 400,
+        headers: { 'Content-Type': 'text/plain' },
+      });
+    }
+
+    // Get unique packages to fetch
+    const uniquePackages = [
+      ...new Map(
+        collection.tools.map((ct) => [
+          ct.tool.package.npmPackageName,
+          {
+            name: ct.tool.package.npmPackageName,
+            version: ct.tool.package.npmVersion,
+          },
+        ])
+      ).values(),
+    ];
+
+    // Fetch package sources (limit to 5 packages for performance)
+    const packageSources = await fetchMultiplePackageSources(uniquePackages.slice(0, 5));
+
+    // Build MCP URLs
+    const baseUrl = process.env.NEXT_PUBLIC_APP_URL || 'https://tpmjs.com';
+    const mcpUrls = {
+      http: `${baseUrl}/api/mcp/${username}/${slug}/http`,
+      sse: `${baseUrl}/api/mcp/${username}/${slug}/sse`,
+    };
+
+    // Generate skills markdown
+    const skillsMarkdown = await generateSkillsMarkdown(
+      {
+        id: collection.id,
+        name: collection.name,
+        slug: collection.slug || slug,
+        description: collection.description,
+        username: user.username,
+      },
+      tools,
+      packageSources,
+      mcpUrls
+    );
+
+    // Save to database for caching
+    await prisma.collection.update({
+      where: { id: collection.id },
+      data: {
+        skillsMarkdown,
+        skillsGeneratedAt: new Date(),
+      },
+    });
+
+    return new Response(skillsMarkdown, {
+      status: 200,
+      headers: {
+        'Content-Type': 'text/markdown; charset=utf-8',
+        'Cache-Control': `public, s-maxage=${CACHE_DURATION_SECONDS}, stale-while-revalidate=86400`,
+        'X-Cache': 'MISS',
+      },
+    });
+  } catch (error) {
+    console.error('[Skills.md Error] GET /:username/collections/:slug/skills.md:', error);
+
+    const message =
+      error instanceof Error ? error.message : 'Failed to generate skills documentation';
+
+    return new Response(`Error: ${message}`, {
+      status: 500,
+      headers: { 'Content-Type': 'text/plain' },
+    });
+  }
+}
diff --git a/apps/web/src/lib/ai/package-source-fetcher.ts b/apps/web/src/lib/ai/package-source-fetcher.ts
new file mode 100644
index 0000000..0e14321
--- /dev/null
+++ b/apps/web/src/lib/ai/package-source-fetcher.ts
@@ -0,0 +1,333 @@
+/**
+ * Package Source Fetcher
+ * Fetches full source code for npm packages via multiple CDN sources
+ */
+
+export interface PackageFile {
+  path: string;
+  content: string;
+}
+
+export interface PackageSource {
+  packageName: string;
+  version: string;
+  files: PackageFile[];
+  packageJson: Record<string, unknown>;
+}
+
+// Maximum files to fetch per package
+const MAX_FILES_PER_PACKAGE = 10;
+// Maximum file size to fetch (50KB)
+const MAX_FILE_SIZE = 50 * 1024;
+// Timeout for individual fetches
+const FETCH_TIMEOUT_MS = 10000;
+
+/**
+ * Fetch with timeout
+ */
+async function fetchWithTimeout(
+  url: string,
+  timeoutMs: number = FETCH_TIMEOUT_MS
+): Promise<Response> {
+  const controller = new AbortController();
+  const timeoutId = setTimeout(() => controller.abort(), timeoutMs);
+
+  try {
+    const response = await fetch(url, {
+      signal: controller.signal,
+      headers: {
+        'User-Agent': 'tpmjs-skills-generator/1.0',
+      },
+    });
+    return response;
+  } finally {
+    clearTimeout(timeoutId);
+  }
+}
+
+/**
+ * Fetch package.json from npm registry
+ */
+async function fetchPackageJson(
+  packageName: string,
+  version?: string
+): Promise<{ packageJson: Record<string, unknown>; resolvedVersion: string }> {
+  const registryUrl = `https://registry.npmjs.org/${packageName}`;
+  const response = await fetchWithTimeout(registryUrl);
+
+  if (!response.ok) {
+    throw new Error(`Failed to fetch package.json for ${packageName}: ${response.status}`);
+  }
+
+  const data = await response.json();
+  const resolvedVersion = version || data['dist-tags']?.latest;
+
+  if (!resolvedVersion) {
+    throw new Error(`Could not resolve version for ${packageName}`);
+  }
+
+  const versionData = data.versions?.[resolvedVersion];
+  if (!versionData) {
+    throw new Error(`Version ${resolvedVersion} not found for ${packageName}`);
+  }
+
+  return { packageJson: versionData, resolvedVersion };
+}
+
+/**
+ * Get list of entry point files from package.json
+ */
+function getEntryPoints(packageJson: Record<string, unknown>): string[] {
+  const entryPoints: string[] = [];
+
+  // Check main entry point
+  if (typeof packageJson.main === 'string') {
+    entryPoints.push(packageJson.main);
+  }
+
+  // Check module entry point (ES modules)
+  if (typeof packageJson.module === 'string') {
+    entryPoints.push(packageJson.module);
+  }
+
+  // Check exports field (modern packages)
+  if (packageJson.exports) {
+    const exports = packageJson.exports as Record<string, unknown>;
+    extractExportPaths(exports, entryPoints);
+  }
+
+  // Check types entry point
+  if (typeof packageJson.types === 'string') {
+    entryPoints.push(packageJson.types);
+  }
+  if (typeof packageJson.typings === 'string') {
+    entryPoints.push(packageJson.typings);
+  }
+
+  // Default fallbacks if nothing found
+  if (entryPoints.length === 0) {
+    entryPoints.push('index.js', 'index.ts', 'src/index.js', 'src/index.ts');
+  }
+
+  // Dedupe and normalize paths
+  const normalized = [...new Set(entryPoints)].map((p) => (p.startsWith('./') ? p.slice(2) : p));
+
+  return normalized;
+}
+
+/**
+ * Recursively extract paths from exports field
+ */
+function extractExportPaths(exports: Record<string, unknown>, paths: string[]): void {
+  for (const value of Object.values(exports)) {
+    if (typeof value === 'string') {
+      paths.push(value);
+    } else if (typeof value === 'object' && value !== null) {
+      // Handle conditional exports like { "import": "./dist/index.mjs", "require": "./dist/index.cjs" }
+      extractExportPaths(value as Record<string, unknown>, paths);
+    }
+  }
+}
+
+/**
+ * Try to fetch a file from unpkg CDN
+ */
+async function fetchFromUnpkg(
+  packageName: string,
+  version: string,
+  filePath: string
+): Promise<string | null> {
+  const url = `https://unpkg.com/${packageName}@${version}/${filePath}`;
+
+  try {
+    const response = await fetchWithTimeout(url);
+    if (!response.ok) return null;
+
+    // Check content length before reading
+    const contentLength = response.headers.get('content-length');
+    if (contentLength && parseInt(contentLength, 10) > MAX_FILE_SIZE) {
+      console.warn(`File too large: ${filePath} (${contentLength} bytes)`);
+      return null;
+    }
+
+    const text = await response.text();
+    if (text.length > MAX_FILE_SIZE) {
+      return null;
+    }
+
+    return text;
+  } catch {
+    return null;
+  }
+}
+
+/**
+ * Try to fetch a file from esm.sh CDN (with ?raw for source)
+ */
+async function fetchFromEsmSh(
+  packageName: string,
+  version: string,
+  filePath: string
+): Promise<string | null> {
+  // esm.sh requires different URL patterns for raw source
+  const url = `https://esm.sh/${packageName}@${version}/${filePath}?raw`;
+
+  try {
+    const response = await fetchWithTimeout(url);
+    if (!response.ok) return null;
+
+    const text = await response.text();
+    if (text.length > MAX_FILE_SIZE) {
+      return null;
+    }
+
+    return text;
+  } catch {
+    return null;
+  }
+}
+
+/**
+ * Fetch a single file from CDN (tries multiple sources)
+ */
+async function fetchFile(
+  packageName: string,
+  version: string,
+  filePath: string
+): Promise<string | null> {
+  // Try unpkg first (more reliable for source files)
+  let content = await fetchFromUnpkg(packageName, version, filePath);
+  if (content) return content;
+
+  // Try esm.sh as fallback
+  content = await fetchFromEsmSh(packageName, version, filePath);
+  if (content) return content;
+
+  return null;
+}
+
+/**
+ * Resolve a relative import path to an absolute path
+ */
+function resolveImportPath(importPath: string, baseDir: string): string {
+  if (importPath.startsWith('./')) {
+    return baseDir ? `${baseDir}/${importPath.slice(2)}` : importPath.slice(2);
+  }
+
+  if (importPath.startsWith('../')) {
+    const parts = baseDir.split('/');
+    const importParts = importPath.split('/');
+    const upCount = importParts.filter((p) => p === '..').length;
+    const remaining = importParts.slice(upCount).join('/');
+    const baseParts = parts.slice(0, parts.length - upCount);
+    return baseParts.length > 0 ? `${baseParts.join('/')}/${remaining}` : remaining;
+  }
+
+  return importPath;
+}
+
+/**
+ * Add file path variations with common extensions
+ */
+function addPathVariations(resolvedPath: string, imports: string[]): void {
+  if (/\.(ts|tsx|js|jsx|mjs|cjs)$/.test(resolvedPath)) {
+    imports.push(resolvedPath);
+  } else {
+    imports.push(`${resolvedPath}.ts`);
+    imports.push(`${resolvedPath}.js`);
+    imports.push(`${resolvedPath}/index.ts`);
+    imports.push(`${resolvedPath}/index.js`);
+  }
+}
+
+/**
+ * Discover additional source files from a fetched file
+ */
+function discoverImports(content: string, basePath: string): string[] {
+  const imports: string[] = [];
+  const baseDir = basePath.includes('/') ? basePath.substring(0, basePath.lastIndexOf('/')) : '';
+
+  // Match import/export statements for local files
+  const importRegex = /(?:import|export).*from\s+['"](\.[^'"]+)['"]/g;
+  const matches = content.matchAll(importRegex);
+
+  for (const match of matches) {
+    const importPath = match[1];
+    if (!importPath) continue;
+
+    const resolvedPath = resolveImportPath(importPath, baseDir);
+    addPathVariations(resolvedPath, imports);
+  }
+
+  return imports;
+}
+
+/**
+ * Fetch source files for an npm package
+ */
+export async function fetchPackageSource(
+  packageName: string,
+  version?: string
+): Promise<PackageSource> {
+  // 1. Fetch package.json to get metadata and entry points
+  const { packageJson, resolvedVersion } = await fetchPackageJson(packageName, version);
+
+  // 2. Get entry points to fetch
+  const entryPoints = getEntryPoints(packageJson);
+  const filesToFetch = new Set<string>();
+  const fetchedFiles: PackageFile[] = [];
+  const fetchedPaths = new Set<string>();
+
+  // Add entry points to fetch queue
+  for (const entry of entryPoints) {
+    filesToFetch.add(entry);
+  }
+
+  // 3. Fetch files breadth-first, discovering imports
+  const queue = [...filesToFetch];
+
+  while (queue.length > 0 && fetchedFiles.length < MAX_FILES_PER_PACKAGE) {
+    const filePath = queue.shift();
+    if (!filePath) break;
+
+    // Skip if already fetched
+    if (fetchedPaths.has(filePath)) continue;
+    fetchedPaths.add(filePath);
+
+    // Try to fetch the file
+    const content = await fetchFile(packageName, resolvedVersion, filePath);
+    if (!content) continue;
+
+    fetchedFiles.push({ path: filePath, content });
+
+    // Discover imports and add to queue
+    const imports = discoverImports(content, filePath);
+    for (const importPath of imports) {
+      if (!fetchedPaths.has(importPath) && !queue.includes(importPath)) {
+        queue.push(importPath);
+      }
+    }
+  }
+
+  return {
+    packageName,
+    version: resolvedVersion,
+    files: fetchedFiles,
+    packageJson,
+  };
+}
+
+/**
+ * Fetch sources for multiple packages concurrently
+ */
+export async function fetchMultiplePackageSources(
+  packages: Array<{ name: string; version?: string }>
+): Promise<PackageSource[]> {
+  const results = await Promise.allSettled(
+    packages.map(({ name, version }) => fetchPackageSource(name, version))
+  );
+
+  return results
+    .filter((r): r is PromiseFulfilledResult<PackageSource> => r.status === 'fulfilled')
+    .map((r) => r.value);
+}
diff --git a/apps/web/src/lib/ai/skills-generator.ts b/apps/web/src/lib/ai/skills-generator.ts
new file mode 100644
index 0000000..0dd1386
--- /dev/null
+++ b/apps/web/src/lib/ai/skills-generator.ts
@@ -0,0 +1,299 @@
+/**
+ * AI-powered skills documentation generator for tool collections
+ * Uses Vercel AI SDK to analyze package source code and generate comprehensive skills.md
+ */
+
+import { openai } from '@ai-sdk/openai';
+import { generateText } from 'ai';
+
+import type { PackageSource } from './package-source-fetcher';
+
+export interface CollectionData {
+  id: string;
+  name: string;
+  slug: string;
+  description: string | null;
+  username: string;
+}
+
+export interface ToolData {
+  id: string;
+  name: string;
+  description: string;
+  packageName: string;
+  inputSchema: unknown | null;
+}
+
+export interface McpUrls {
+  http: string;
+  sse: string;
+}
+
+const SKILLS_SYSTEM_PROMPT = `You are an expert technical writer generating machine-consumable skills documentation for AI agents.
+
+Your task is to analyze actual source code from npm packages and generate a comprehensive skills.md document that serves as a capability contract for AI agents like Claude Code, MCP clients, and tool routers.
+
+Guidelines:
+1. **Accuracy First**: Only document capabilities you can verify from the source code
+2. **Real Examples**: Generate code examples based on actual function signatures and patterns in the source
+3. **Input/Output Schemas**: Extract exact TypeScript types when available
+4. **Source Analysis**: Provide insights about implementation details that help agents use tools correctly
+5. **Integration Focus**: Show multiple ways to integrate (ESM, NPM, CLI, REST API, MCP)
+6. **Constraints**: Document what tools can and cannot do based on code analysis
+
+Format the output as clean, well-organized Markdown with proper code blocks and syntax highlighting.`;
+
+function buildSkillsPrompt(
+  collection: CollectionData,
+  tools: ToolData[],
+  packageSources: PackageSource[],
+  mcpUrls: McpUrls
+): string {
+  const baseUrl = mcpUrls.http.replace(/\/api\/mcp\/.*$/, '');
+  const collectionUrl = `${baseUrl}/${collection.username}/collections/${collection.slug}`;
+
+  // Build package source context
+  const sourceContext = packageSources
+    .map((pkg) => {
+      const filesContent = pkg.files
+        .map((f) => `### ${f.path}\n\`\`\`typescript\n${f.content.slice(0, 5000)}\n\`\`\``)
+        .join('\n\n');
+
+      return `## Package: ${pkg.packageName}@${pkg.version}
+
+**package.json exports:**
+\`\`\`json
+${JSON.stringify(
+  {
+    main: pkg.packageJson.main,
+    module: pkg.packageJson.module,
+    exports: pkg.packageJson.exports,
+    types: pkg.packageJson.types,
+  },
+  null,
+  2
+)}
+\`\`\`
+
+**Source Files:**
+${filesContent}`;
+    })
+    .join('\n\n---\n\n');
+
+  // Build tools context
+  const toolsContext = tools
+    .map(
+      (t) => `- **${t.name}** (${t.packageName})
+  Description: ${t.description}
+  Input Schema: ${t.inputSchema ? JSON.stringify(t.inputSchema, null, 2) : 'Not available'}`
+    )
+    .join('\n\n');
+
+  return `Generate a skills.md document for this collection.
+
+## Collection Info
+- **Name:** ${collection.name}
+- **Owner:** @${collection.username}
+- **Slug:** ${collection.slug}
+- **Description:** ${collection.description || 'No description'}
+- **Collection URL:** ${collectionUrl}
+
+## MCP Endpoints
+- **HTTP:** ${mcpUrls.http}
+- **SSE:** ${mcpUrls.sse}
+
+## Tools in Collection (${tools.length} total)
+${toolsContext}
+
+## Package Source Code
+Analyze the following source code to understand tool implementations:
+
+${sourceContext}
+
+---
+
+Generate the skills.md document following this EXACT structure:
+
+# Agent Skills Declaration: ${collection.name}
+
+> Machine-consumable capability contract for AI agents (Claude Code, MCP clients, tool routers)
+
+## 1. Agent Identity
+
+**Collection:** ${collection.name}
+**Owner:** @${collection.username}
+**Primary Role:** [Derive from tools and description]
+**Tool Count:** ${tools.length} tools from ${new Set(tools.map((t) => t.packageName)).size} packages
+
+---
+
+## 2. Core Skills
+
+[For each tool, include:]
+### Skill: [tool.name]
+**Package:** \`[package.npmPackageName]\`
+**Description:** [tool.description]
+
+**Input Schema:**
+\`\`\`typescript
+[Extract from source code]
+\`\`\`
+
+**Output Format:** [Analyze from source]
+
+**Source Analysis:**
+[AI-generated insights from reading actual code - what does this tool actually do internally?]
+
+**Example Usage:**
+\`\`\`typescript
+[Real example based on source code patterns]
+\`\`\`
+
+---
+
+## 3. Direct Usage (Without MCP/CLI/REST)
+
+### ESM Import (Browser/Deno)
+\`\`\`typescript
+[Generate real ESM import examples using esm.sh]
+\`\`\`
+
+### NPM Import (Node.js)
+\`\`\`typescript
+[Generate real NPM import examples]
+\`\`\`
+
+### Multi-Tool Workflow
+\`\`\`typescript
+[Generate a workflow example combining multiple tools]
+\`\`\`
+
+---
+
+## 4. Agent Integration Methods
+
+### 4.1 CLI (\`tpm run\`)
+
+Install CLI:
+\`\`\`bash
+npm install -g @tpmjs/cli
+\`\`\`
+
+Run a tool:
+\`\`\`bash
+tpm run --collection ${collection.username}/${collection.slug} --tool [tool-name] --args '{"key": "value"}'
+\`\`\`
+
+With environment variables:
+\`\`\`bash
+SOME_API_KEY=xxx tpm run --collection ${collection.username}/${collection.slug} --tool [tool-name]
+\`\`\`
+
+### 4.2 REST API (JSON-RPC 2.0)
+
+**Endpoint:** \`POST ${mcpUrls.http}\`
+
+\`\`\`bash
+curl -X POST ${mcpUrls.http} \\
+  -H "Content-Type: application/json" \\
+  -H "Authorization: Bearer YOUR_API_KEY" \\
+  -d '{
+    "jsonrpc": "2.0",
+    "id": 1,
+    "method": "tools/call",
+    "params": {
+      "name": "[mcp-tool-name]",
+      "arguments": { ... }
+    }
+  }'
+\`\`\`
+
+**List available tools:**
+\`\`\`bash
+curl -X POST ${mcpUrls.http} \\
+  -H "Content-Type: application/json" \\
+  -H "Authorization: Bearer YOUR_API_KEY" \\
+  -d '{"jsonrpc": "2.0", "id": 1, "method": "tools/list"}'
+\`\`\`
+
+### 4.3 MCP Server
+
+**URLs:**
+- HTTP: \`${mcpUrls.http}\`
+- SSE: \`${mcpUrls.sse}\`
+
+**Claude Desktop Config (\`claude_desktop_config.json\`):**
+\`\`\`json
+{
+  "mcpServers": {
+    "${collection.slug}": {
+      "url": "${mcpUrls.sse}",
+      "headers": {
+        "Authorization": "Bearer YOUR_API_KEY"
+      }
+    }
+  }
+}
+\`\`\`
+
+**Local MCP Server (via CLI):**
+\`\`\`bash
+tpm mcp serve ${collection.username}/${collection.slug}
+\`\`\`
+
+---
+
+## 5. Constraints & Safety
+
+**This collection does NOT:**
+[Generate based on tool analysis - what are the limitations?]
+
+**Rate Limits:**
+- Free tier: 100 requests/hour
+- Authenticated: 1000 requests/hour
+
+**Authentication:**
+- Public collections: No auth required for read
+- Tool execution: API key required
+
+---
+
+## 6. Versioning
+
+**Skills Version:** 1.0.0
+**Generated:** [Current timestamp]
+**Collection Updated:** [Will be filled in]
+
+---
+
+## 7. Canonical References
+
+- TPMJS: https://tpmjs.com
+- MCP Protocol: https://modelcontextprotocol.io
+- Collection Page: ${collectionUrl}`;
+}
+
+/**
+ * Generate skills.md markdown for a collection
+ */
+export async function generateSkillsMarkdown(
+  collection: CollectionData,
+  tools: ToolData[],
+  packageSources: PackageSource[],
+  mcpUrls: McpUrls
+): Promise<string> {
+  if (tools.length === 0) {
+    throw new Error('Collection must have at least one tool to generate skills documentation');
+  }
+
+  const prompt = buildSkillsPrompt(collection, tools, packageSources, mcpUrls);
+
+  const { text } = await generateText({
+    model: openai('gpt-4.1-mini'),
+    system: SKILLS_SYSTEM_PROMPT,
+    prompt,
+    temperature: 0.3, // Lower temperature for more factual output
+  });
+
+  return text;
+}
diff --git a/docs/PRD-omega.md b/docs/PRD-omega.md
new file mode 100644
index 0000000..309c75b
--- /dev/null
+++ b/docs/PRD-omega.md
@@ -0,0 +1,287 @@
+# Omega PRD (Draft v0.1)
+
+Project name: Omega  
+Path: `/omega`  
+Owner: TPMJS  
+Status: Draft
+
+## 1) Executive Summary
+Omega is a TPMJS side project that demonstrates what happens when a single agent can discover and execute tools from the entire TPM registry. The vision is playful and bold: "the last robot/AI you ever need" (Skynet energy, but friendly). Omega will ship as a dedicated chat experience at `/omega` and a public REST API that lets anyone start and continue conversations. The product must highlight tool discovery, selection, and execution at scale, while handling missing credentials gracefully.
+
+## 2) Problem Statement
+TPMJS has powerful registry and tool execution capabilities, but the end-user experience does not yet show how these components combine into a single, continuous, tool-driven agent. Users need a simple way to:
+- Ask for outcomes, not tools
+- Watch the agent discover the right tool(s)
+- Provide credentials only when needed
+- Continue execution without restarting
+
+## 3) Vision
+Omega is a single agent with access to the TPM registry, able to choose and run the right tools to complete tasks. Users can chat or call an API and see a transparent, step-by-step tool workflow. It is a public showcase that can become a standalone product over time.
+
+## 4) Goals and Success Metrics
+### Goals
+- Showcase tool discovery and execution at scale (1M+ tools).
+- Deliver a great developer experience (DX) with simple API endpoints.
+- Make credential handling safe, clear, and resumable.
+- Provide a polished chat UI that is fast and transparent.
+
+### Success Metrics (Early)
+- Conversations created per week
+- % conversations that use at least one tool
+- Median time to first tool call
+- Completion rate (user gets a result without abandoning)
+- Avg. tool-call count per conversation
+- User feedback: "felt like a real agent"
+
+## 5) Non-Goals (v0)
+- Multi-agent orchestration
+- Enterprise SSO or complex permission tiers
+- Guaranteed deterministic tool selection
+- Complex billing or metering
+
+## 6) Target Users and Use Cases
+### Personas
+- Developers evaluating TPMJS or building on it (primary for v0)
+- Power users exploring tool chains
+- Curious users experimenting with AI + tools
+
+### Example Use Cases
+- "Summarize a website and turn it into a checklist"
+- "Analyze a CSV and generate insights"
+- "Generate a proposal and send it via email"
+- "Search for a tool to do X and run it"
+
+## 7) Product Requirements
+
+### 7.1 Chat UI (`/omega`)
+- Landing intro before first chat
+- New conversation
+- Message stream with agent responses
+- Tool activity panel (tool name, inputs, outputs, status)
+- Debug mode (raw JSON) and "view raw" toggles for tool I/O
+- Shareable conversation URL
+- Requires login (v0)
+- Credential prompt when missing env vars
+- Resume execution after credentials are provided
+- Planning panel for proposed plans (only when requested or task is complex)
+- Cancel running tool execution or full conversation run
+- Multi-user conversations: shared links allow anyone with URL to send messages
+- Show participant display names and Gravatar avatars (email-based)
+- Export conversation to JSON and Markdown
+- Lazy-load long outputs and message history (Virtuoso-based)
+- Auto-generate conversation title after first response
+- System status banner (registry + executor health)
+- Empty state includes creative sample prompts
+- Light/dark theme supported (use TPMJS style guide)
+
+### 7.2 REST API
+Endpoints (initial):
+- `POST /api/omega/conversations`  
+  Create a conversation. Returns conversation ID and optional title.
+- `POST /api/omega/conversations/:id/messages`  
+  Send a message and stream response (SSE).
+- `GET /api/omega/conversations/:id`  
+  Fetch conversation history.
+- `POST /api/omega/conversations/:id/credentials`  
+  Provide missing env vars and resume.
+- `POST /api/omega/conversations/:id/plan`  
+  Optional: generate plan(s) without execution.
+- API access requires TPMJS API key (standard user keys).
+- Shared conversations accept either logged-in session or TPMJS API key.
+
+### 7.3 Tool Discovery and Execution
+- Agent can search registry and execute tools.
+- Default path: registry search + execute tools (entire registry).
+- Registry execution uses TPMJS sandbox executor.
+- Allow tool usage transparency in UI and API responses.
+- Allow users to pin preferred tools for the agent to prioritize.
+- Pinned tools are always included in the plan/tool set.
+- Max tool calls per message: 200 (AI SDK `maxSteps`).
+
+### 7.6 System Prompt and Safety Controls
+- Omega ships with a comprehensive default system prompt.
+- Users can extend/override with a custom prompt in Omega settings.
+- Safe Mode toggle: require explicit approval before tool execution (human-in-the-loop).
+
+### 7.4 Planning
+- Agent can propose one or more plans before running tools.
+- Plans show tool order, purpose, expected outputs, and confidence.
+- User can select a plan or ask for another.
+- Planning can be auto-run for complex tasks or when explicitly requested.
+
+### 7.5 Credentials / Missing Env Vars
+- Detect missing env vars for tools.
+- Pause execution when required env vars missing.
+- Provide a credential prompt in UI (inline modal).
+- Resume execution after credentials are supplied.
+- Allow API callers to supply env vars in requests.
+- Store credentials encrypted per-user in Omega settings.
+
+## 8) UX Flows
+
+### Flow A: Standard Chat
+1. User sends a message.
+2. Agent responds, streams tokens.
+3. Tool calls appear in tool panel.
+4. Tool results appear inline.
+5. Conversation persists.
+
+### Flow B: Missing Credentials
+1. Tool requires env vars.
+2. Agent pauses and returns `MISSING_ENV_VARS`.
+3. UI displays required keys and descriptions.
+4. User provides keys.
+5. Execution resumes from same step.
+
+### Flow C: Planning
+1. User sends a request.
+2. Agent responds with one or more plans.
+3. User selects a plan or asks for changes.
+4. Execution starts.
+
+### Flow D: Tool Failure + Alternatives
+1. Tool fails or missing credentials.
+2. Agent searches for alternative tools.
+3. Agent presents options and required keys in a table.
+4. User selects a tool; execution resumes.
+
+## 9) Data Model (High-Level)
+- Conversation: id, userId (optional), createdAt, updatedAt, title
+- Message: id, conversationId, role, content, createdAt
+- ToolRun: id, conversationId, messageId, toolId, input, output, status, duration
+- Plan: id, conversationId, messageId, steps, selectedPlanId, createdAt
+- CredentialRequest: id, conversationId, toolRunId, requiredKeys, status, createdAt
+- CredentialStore: userId, keyName, encryptedValue, scope (global/tool)
+
+## 10) Security and Privacy
+- Encrypt stored credentials at rest.
+- Never expose full secret values in UI or logs.
+- Redact secrets from tool outputs when needed.
+- Rate limit API endpoints to prevent abuse.
+- Log rate limit events (even if limits are effectively unlimited at launch).
+- Audit logs for tool calls and credential access.
+
+## 11) Technical Foundations (Reuse Existing)
+From current codebase:
+- Chat + SSE (agent conversations, tool-call streaming, message persistence):
+  `apps/web/src/app/api/agents/[id]/conversation/[conversationId]/route.ts`
+- Pretty URL variant for agent chat (API key scopes + owner/public checks):
+  `apps/web/src/app/api/[username]/agents/[agentSlug]/conversation/[conversationId]/route.ts`
+- Chat UI with tool panel + streaming + Virtuoso pagination:
+  `apps/web/src/app/agents/[id]/chat/[chatId]/page.tsx`
+- Tool execution wrapper + AI SDK tool definition builder:
+  `apps/web/src/lib/ai-agent/tool-executor-agent.ts`
+- Executor resolution + sandbox execution:
+  `apps/web/src/lib/executors/index.ts`
+- Registry meta-tools for search + execute:
+  `packages/tools/registrySearch`, `packages/tools/registryExecute`
+- Env var detection and missing var reporting:
+  `apps/web/src/lib/agents/env-helpers.ts`
+- Env var cascade and caller-provided env support:
+  `apps/web/src/lib/agents/build-tools.ts`
+- MCP handler with env-var validation for non-owners:
+  `apps/web/src/lib/mcp/handlers.ts`
+- Credential encryption utilities and API key storage:
+  `apps/web/src/lib/crypto/api-keys.ts`, `apps/web/src/app/api/user/api-keys/route.ts`
+- Existing planning tool for generatePlans:
+  `packages/tools/official/tool-selection-plan`
+
+## 12) Product Decisions (Initial)
+- Primary audience: developers.
+- Auth: login required (v0).
+- Tool access: entire registry via registry search + execute.
+- Model: default `gpt-4.1-mini`; users choose provider/model if they have keys.
+- Planning: only on explicit request or when task complexity is high.
+- Credentials: stored per-user, scoped to Omega, managed in Omega settings.
+- Safety: no additional tool restrictions beyond existing TPMJS policies.
+- Sharing: anyone with conversation URL can participate.
+- Retention: conversations are stored indefinitely.
+- Tool output: summarized by default with "view raw" toggle.
+- Cancel: users can stop a running tool or full run.
+- Rate limits: system in place, effectively unlimited at launch.
+- Participants: show display names + Gravatar avatars.
+- API auth: requires standard TPMJS API key.
+- Tool selection UX: agent-only (no manual marketplace UI for v0).
+- Files: uploads deferred to a later phase.
+- Streaming: stream tool outputs where supported by the AI SDK.
+- UI: follow TPMJS style guide (no bespoke theme for v0).
+- Conversation IDs: UUIDs.
+- Credentials: per-user only (no team/workspace creds in v0).
+- Conversation context: keep full history (no auto-summarization in v0).
+- Compliance: no special compliance requirements at launch.
+- Tool budget indicator: not required.
+- Shared conversation execution: only owner-stored credentials may run tools.
+- Large outputs: lazy-load via virtualization.
+- Export: JSON and Markdown supported.
+- Memory: per-conversation only.
+- Conversation titles: auto-generated after first response (AI SDK).
+- Reactions: not required for v0.
+- Tool reasoning: show only when model returns reasoning tokens.
+- Deletion: users cannot delete conversations in v0.
+- API keys: use existing TPMJS API keys only.
+- Discovery: conversations are not listed; URL-only access.
+- Context control: no reset button, use new chat instead.
+- Tool allowlist: no allowlist (pins only).
+- Audit/log retention: store all tool I/O indefinitely.
+- Shared chat auth: accept session auth or TPMJS API key.
+- Landing: intro view before first chat.
+- Empty state: creative sample prompts.
+- Theme: light/dark via style guide.
+
+## 13) API Response Contracts (Draft)
+### Missing Env Vars
+```json
+{
+  "success": false,
+  "error": "Missing required environment variables",
+  "details": {
+    "code": "MISSING_ENV_VARS",
+    "missingVars": [
+      { "name": "FIRECRAWL_API_KEY", "description": "API key for Firecrawl" }
+    ],
+    "hint": "Provide these variables in the 'env' field of your request"
+  }
+}
+```
+
+### Plan Response
+```json
+{
+  "plan": [
+    { "stepNumber": 1, "toolName": "registrySearchTool", "purpose": "Find tools for web scraping" },
+    { "stepNumber": 2, "toolName": "registryExecuteTool", "purpose": "Run the chosen tool" }
+  ],
+  "confidence": 0.74,
+  "reasoning": "Found 2 relevant tools for this task"
+}
+```
+
+## 14) Performance and Reliability
+- SSE streams should stay responsive under load.
+- Tool execution timeouts (default 5 minutes).
+- Concurrency controls per conversation.
+- Circuit breaker for broken tools or executors.
+
+## 15) Observability
+- Log tool calls and errors.
+- Track token usage per conversation.
+- Surface tool success/failure rates.
+- Store conversation logs for debugging.
+
+## 16) Risks and Mitigations
+- Tool quality variance: provide confidence and fallbacks.
+- Missing credentials: pause/resume flow with clear UX.
+- Abuse risk: rate limiting + auth.
+- Tool execution cost: usage tracking + limits.
+
+## 17) Rollout Plan
+1. Internal alpha (team-only)
+2. Public beta with rate limits
+3. Full public launch
+
+## 18) Open Questions
+- System prompt ownership and defaults (versioning and updates)?
+- Safe Mode default on/off?
+- How many plans should Omega propose by default?
+ - Webhook support for API events (future).
+ - Profiles for system prompts (future).
diff --git a/packages/cli/README.md b/packages/cli/README.md
index 9a60d56..cb6c3f2 100644
--- a/packages/cli/README.md
+++ b/packages/cli/README.md
@@ -72,6 +72,24 @@ tpm collection add       # Add tools to a collection
 tpm collection remove    # Remove tools from a collection
 ```
 
+### Run Tools
+
+Execute tools directly from collections via the MCP API:
+
+```bash
+# Basic usage
+tpm run -c ajax/unsandbox -t execute --args '{"code":"print(1)","language":"python"}'
+
+# With environment variables
+tpm run -c ajax/my-collection -t myTool --env API_KEY=xxx --env DEBUG=true
+
+# Using process environment
+OPENAI_API_KEY=xxx tpm run -c ajax/ai-tools -t generate
+
+# JSON output for scripting
+tpm run -c ajax/tools -t base64Encode --args '{"data":"hello"}' --json | jq .result
+```
+
 ### MCP Integration
 
 ```bash
diff --git a/packages/cli/src/commands/run.ts b/packages/cli/src/commands/run.ts
new file mode 100644
index 0000000..c632240
--- /dev/null
+++ b/packages/cli/src/commands/run.ts
@@ -0,0 +1,321 @@
+import { Command, Flags } from '@oclif/core';
+
+import { getApiKey, getApiUrl } from '../lib/config.js';
+import { createOutput, type OutputFormatter } from '../lib/output.js';
+
+interface JsonRpcResponse {
+  jsonrpc: '2.0';
+  id: string | number | null;
+  result?: {
+    content?: Array<{ type: string; text: string }>;
+    [key: string]: unknown;
+  };
+  error?: {
+    code: number;
+    message: string;
+    data?: unknown;
+  };
+}
+
+interface RunFlags {
+  collection: string;
+  tool: string;
+  args: string;
+  env?: string[];
+  json: boolean;
+  verbose: boolean;
+  timeout: number;
+}
+
+/**
+ * Parse and validate collection identifier
+ */
+function parseCollection(collection: string): { username: string; slug: string } | null {
+  const parts = collection.split('/');
+  if (parts.length !== 2) return null;
+  const [username, slug] = parts;
+  if (!username || !slug) return null;
+  return { username, slug };
+}
+
+/**
+ * Parse JSON tool arguments
+ */
+function parseToolArgs(argsStr: string): Record<string, unknown> | null {
+  try {
+    return JSON.parse(argsStr) as Record<string, unknown>;
+  } catch {
+    return null;
+  }
+}
+
+/**
+ * Build environment variables from process.env and explicit flags
+ */
+function buildEnvVars(envFlags: string[] | undefined, output: OutputFormatter): Record<string, string> {
+  const envVars: Record<string, string> = {};
+
+  // Add process env vars that might be relevant
+  for (const [key, value] of Object.entries(process.env)) {
+    if (value && (key.endsWith('_API_KEY') || key.endsWith('_TOKEN') || key.endsWith('_SECRET'))) {
+      envVars[key] = value;
+    }
+  }
+
+  // Add explicitly passed env vars (override process env)
+  if (envFlags) {
+    for (const envPair of envFlags) {
+      const [key, ...valueParts] = envPair.split('=');
+      if (key && valueParts.length > 0) {
+        envVars[key] = valueParts.join('=');
+      } else {
+        output.warning(`Invalid env var format: ${envPair}. Use: KEY=value`);
+      }
+    }
+  }
+
+  return envVars;
+}
+
+/**
+ * Display verbose request info
+ */
+function displayVerboseInfo(
+  output: OutputFormatter,
+  mcpUrl: string,
+  tool: string,
+  toolArgs: Record<string, unknown>,
+  envVars: Record<string, string>
+): void {
+  output.info(`MCP Endpoint: ${mcpUrl}`);
+  output.info(`Tool: ${tool}`);
+  output.info(`Arguments: ${JSON.stringify(toolArgs, null, 2)}`);
+  if (Object.keys(envVars).length > 0) {
+    output.info(`Environment: ${Object.keys(envVars).join(', ')}`);
+  }
+  output.divider();
+}
+
+/**
+ * Handle HTTP error response
+ */
+async function handleHttpError(response: Response, output: OutputFormatter): Promise<void> {
+  const errorText = await response.text();
+  try {
+    const errorJson = JSON.parse(errorText) as { error?: { message?: string }; message?: string };
+    output.error(errorJson.error?.message || errorJson.message || `HTTP ${response.status}`);
+  } catch {
+    output.error(`HTTP ${response.status}: ${errorText}`);
+  }
+}
+
+/**
+ * Display MCP response content
+ */
+function displayContent(result: JsonRpcResponse['result'], output: OutputFormatter): void {
+  if (!result) {
+    output.text('(empty result)');
+    return;
+  }
+
+  if (result.content && Array.isArray(result.content)) {
+    for (const item of result.content) {
+      displayContentItem(item, output);
+    }
+  } else {
+    output.text(JSON.stringify(result, null, 2));
+  }
+}
+
+/**
+ * Display a single content item
+ */
+function displayContentItem(item: { type: string; text: string }, output: OutputFormatter): void {
+  if (item.type === 'text') {
+    try {
+      const parsed = JSON.parse(item.text);
+      output.text(JSON.stringify(parsed, null, 2));
+    } catch {
+      output.text(item.text);
+    }
+  } else if (item.type === 'image') {
+    output.info('[Image content]');
+  } else {
+    output.text(JSON.stringify(item, null, 2));
+  }
+}
+
+/**
+ * Handle execution error
+ */
+function handleExecutionError(error: unknown, flags: RunFlags, output: OutputFormatter): void {
+  if (error instanceof Error) {
+    if (error.name === 'AbortError') {
+      output.error(`Request timed out after ${flags.timeout} seconds`);
+    } else {
+      output.error(error.message);
+    }
+  } else {
+    output.error('Unknown error occurred');
+  }
+
+  if (flags.verbose) {
+    output.info(`Full error: ${String(error)}`);
+  }
+}
+
+export default class Run extends Command {
+  static description = 'Execute a tool from a collection via MCP';
+
+  static examples = [
+    '<%= config.bin %> run -c ajax/unsandbox -t execute --args \'{"code":"print(1)","language":"python"}\'',
+    '<%= config.bin %> run --collection ajax/ajax-collection --tool base64Encode --args \'{"data":"hello"}\'',
+    'OPENAI_API_KEY=xxx <%= config.bin %> run -c ajax/my-collection -t myTool',
+    '<%= config.bin %> run -c ajax/tools -t search --args \'{"query":"test"}\' --json',
+    '<%= config.bin %> run -c ajax/tools -t search --env API_KEY=xxx --env DEBUG=true',
+  ];
+
+  static flags = {
+    collection: Flags.string({
+      char: 'c',
+      description: 'Collection identifier (username/slug)',
+      required: true,
+    }),
+    tool: Flags.string({
+      char: 't',
+      description: 'Tool name (MCP format: package--toolName)',
+      required: true,
+    }),
+    args: Flags.string({
+      char: 'a',
+      description: 'Tool arguments as JSON string',
+      default: '{}',
+    }),
+    env: Flags.string({
+      char: 'e',
+      description: 'Environment variables (key=value)',
+      multiple: true,
+    }),
+    json: Flags.boolean({
+      description: 'Output result as JSON',
+      default: false,
+    }),
+    verbose: Flags.boolean({
+      char: 'v',
+      description: 'Show verbose output',
+      default: false,
+    }),
+    timeout: Flags.integer({
+      description: 'Request timeout in seconds',
+      default: 60,
+    }),
+  };
+
+  async run(): Promise<void> {
+    const { flags } = await this.parse(Run);
+    const output = createOutput(flags);
+
+    // Parse and validate inputs
+    const collection = parseCollection(flags.collection);
+    if (!collection) {
+      output.error('Invalid collection format. Use: username/slug');
+      return;
+    }
+
+    const toolArgs = parseToolArgs(flags.args);
+    if (!toolArgs) {
+      output.error('Invalid JSON in --args flag');
+      return;
+    }
+
+    // Build request components
+    const envVars = buildEnvVars(flags.env, output);
+    const baseUrl = getApiUrl().replace(/\/api$/, '');
+    const mcpUrl = `${baseUrl}/api/mcp/${collection.username}/${collection.slug}/http`;
+
+    const apiKey = getApiKey();
+    if (!apiKey) {
+      output.warning('No API key configured. Some operations may fail.');
+      output.warning('Run `tpm auth login` to authenticate.');
+    }
+
+    if (flags.verbose) {
+      displayVerboseInfo(output, mcpUrl, flags.tool, toolArgs, envVars);
+    }
+
+    // Execute the request
+    await this.executeToolCall(flags, output, mcpUrl, apiKey, toolArgs, envVars);
+  }
+
+  private async executeToolCall(
+    flags: RunFlags,
+    output: OutputFormatter,
+    mcpUrl: string,
+    apiKey: string | undefined,
+    toolArgs: Record<string, unknown>,
+    envVars: Record<string, string>
+  ): Promise<void> {
+    const jsonRpcRequest = {
+      jsonrpc: '2.0',
+      id: Date.now(),
+      method: 'tools/call',
+      params: {
+        name: flags.tool,
+        arguments: toolArgs,
+        ...(Object.keys(envVars).length > 0 ? { env: envVars } : {}),
+      },
+    };
+
+    const spinner = output.spinner(`Executing ${flags.tool}...`);
+
+    try {
+      const controller = new AbortController();
+      const timeoutId = setTimeout(() => controller.abort(), flags.timeout * 1000);
+
+      const response = await fetch(mcpUrl, {
+        method: 'POST',
+        headers: {
+          'Content-Type': 'application/json',
+          ...(apiKey ? { Authorization: `Bearer ${apiKey}` } : {}),
+        },
+        body: JSON.stringify(jsonRpcRequest),
+        signal: controller.signal,
+      });
+
+      clearTimeout(timeoutId);
+
+      if (!response.ok) {
+        spinner.fail('Request failed');
+        await handleHttpError(response, output);
+        return;
+      }
+
+      const result = (await response.json()) as JsonRpcResponse;
+      spinner.stop();
+
+      this.handleResult(flags, output, result);
+    } catch (error) {
+      spinner.fail('Execution failed');
+      handleExecutionError(error, flags, output);
+    }
+  }
+
+  private handleResult(flags: RunFlags, output: OutputFormatter, result: JsonRpcResponse): void {
+    if (result.error) {
+      output.error(`Tool execution failed: ${result.error.message}`);
+      if (flags.verbose && result.error.data) {
+        output.info(`Error data: ${JSON.stringify(result.error.data, null, 2)}`);
+      }
+      return;
+    }
+
+    if (flags.json) {
+      output.json(result.result);
+      return;
+    }
+
+    output.success('Tool execution complete');
+    output.divider();
+    displayContent(result.result, output);
+  }
+}
diff --git a/packages/db/prisma/schema.prisma b/packages/db/prisma/schema.prisma
index 03953b9..c5ef34b 100644
--- a/packages/db/prisma/schema.prisma
+++ b/packages/db/prisma/schema.prisma
@@ -458,6 +458,10 @@ model Collection {
   useCases            Json?     @map("use_cases") @db.JsonB
   useCasesGeneratedAt DateTime? @map("use_cases_generated_at")
 
+  // AI-generated skills documentation
+  skillsMarkdown      String?   @map("skills_markdown") @db.Text
+  skillsGeneratedAt   DateTime? @map("skills_generated_at")
+
   // Timestamps
   createdAt   DateTime @default(now()) @map("created_at")
   updatedAt   DateTime @updatedAt @map("updated_at")