deepwiki_fetch
Fetch and convert Deepwiki documentation repositories into Markdown format for easy access and integration.
Instructions
Fetch a deepwiki.com repo and return Markdown
Input Schema
TableJSON Schema
| Name | Required | Description | Default |
|---|---|---|---|
| url | Yes | should be a URL, owner/repo name (e.g. "vercel/ai"), a two-word "owner repo" form (e.g. "vercel ai"), or a single library keyword | |
| maxDepth | No | Can fetch a single site => maxDepth 0 or multiple/all sites => maxDepth 1 | |
| mode | No | aggregate | |
| verbose | No |
Implementation Reference
- src/tools/deepwiki.ts:19-131 (handler)The handler function that normalizes the input URL, performs validation, crawls the deepwiki.com repository using the crawl utility, converts HTML pages to Markdown, and returns the content as an array of text blocks.
async (input) => { // Normalize the URL to support short forms const normalizedInput = { ...input } if (typeof normalizedInput.url === 'string') { let url = normalizedInput.url.trim() // Only transform when it is not already an explicit HTTP(S) URL if (!/^https?:\/\//.test(url)) { // Check if the URL is already in the owner/repo format if (/^[^/]+\/[^/]+$/.test(url)) { // Already in owner/repo format, keep it as is // Just prefix with deepwiki.com } // Single word/term with no slash - process differently else if (/^[^/]+$/.test(url)) { // For single words, first try to extract a meaningful keyword if the input has spaces if (url.includes(' ')) { const extracted = extractKeyword(url) if (extracted) { url = extracted } } // Try to resolve the single term against GitHub try { const repo = await resolveRepo(url) // "owner/repo" url = repo } catch { // Fallback to previous behaviour for backward compatibility url = `defaultuser/${url}` // TODO: replace defaultuser logic } } // Other formats (phrases with slashes that don't match owner/repo) else { // Try to extract a library keyword from a free form phrase const extracted = extractKeyword(url) if (extracted) { // Resolve the extracted keyword try { const repo = await resolveRepo(extracted) url = repo } catch { url = `defaultuser/${extracted}` } } } // At this point url should be "owner/repo" url = `https://deepwiki.com/${url}` } normalizedInput.url = url } const parse = FetchRequest.safeParse(normalizedInput) if (!parse.success) { const err: z.infer<typeof ErrorEnvelope> = { status: 'error', code: 'VALIDATION', message: 'Request failed schema validation', details: parse.error.flatten(), } return err } const req = parse.data const root = new URL(req.url) if (req.maxDepth > 1) { const err: z.infer<typeof ErrorEnvelope> = { status: 'error', code: 'VALIDATION', message: 'maxDepth > 1 is not allowed', } return err } if (root.hostname !== 'deepwiki.com') { const err: z.infer<typeof ErrorEnvelope> = { status: 'error', code: 'DOMAIN_NOT_ALLOWED', message: 'Only deepwiki.com domains are allowed', } return err } // Progress emitter function emitProgress(e: any) { // Progress reporting is not supported in this context because McpServer does not have a sendEvent method. } const crawlResult = await crawl({ root, maxDepth: req.maxDepth, emit: emitProgress, verbose: req.verbose, }) // Convert each page const pages = await Promise.all( Object.entries(crawlResult.html).map(async ([path, html]) => ({ path, markdown: await htmlToMarkdown(html, req.mode), })), ) return { content: pages.map(page => ({ type: 'text', text: `# ${page.path}\n\n${page.markdown}`, })), } }, - src/schemas/deepwiki.ts:9-18 (schema)Zod schema defining the input parameters for the deepwiki_fetch tool: url, maxDepth, mode, and verbose.
export const FetchRequest = z.object({ /** Deepwiki repo URL, eg https://deepwiki.com/user/repo */ url: z.string().describe('should be a URL, owner/repo name (e.g. "vercel/ai"), a two-word "owner repo" form (e.g. "vercel ai"), or a single library keyword'), /** Crawl depth limit: 0 means only the root page */ maxDepth: z.number().int().min(0).max(1).default(1).describe('Can fetch a single site => maxDepth 0 or multiple/all sites => maxDepth 1'), /** Conversion mode */ mode: ModeEnum.default('aggregate'), /** Verbose logging flag */ verbose: z.boolean().default(false), }) - src/tools/deepwiki.ts:14-133 (registration)The deepwikiTool function that registers the 'deepwiki_fetch' tool with the MCP context using mcp.tool, providing name, description, input schema, and handler.
export function deepwikiTool({ mcp }: McpToolContext) { mcp.tool( 'deepwiki_fetch', 'Fetch a deepwiki.com repo and return Markdown', FetchRequest.shape, async (input) => { // Normalize the URL to support short forms const normalizedInput = { ...input } if (typeof normalizedInput.url === 'string') { let url = normalizedInput.url.trim() // Only transform when it is not already an explicit HTTP(S) URL if (!/^https?:\/\//.test(url)) { // Check if the URL is already in the owner/repo format if (/^[^/]+\/[^/]+$/.test(url)) { // Already in owner/repo format, keep it as is // Just prefix with deepwiki.com } // Single word/term with no slash - process differently else if (/^[^/]+$/.test(url)) { // For single words, first try to extract a meaningful keyword if the input has spaces if (url.includes(' ')) { const extracted = extractKeyword(url) if (extracted) { url = extracted } } // Try to resolve the single term against GitHub try { const repo = await resolveRepo(url) // "owner/repo" url = repo } catch { // Fallback to previous behaviour for backward compatibility url = `defaultuser/${url}` // TODO: replace defaultuser logic } } // Other formats (phrases with slashes that don't match owner/repo) else { // Try to extract a library keyword from a free form phrase const extracted = extractKeyword(url) if (extracted) { // Resolve the extracted keyword try { const repo = await resolveRepo(extracted) url = repo } catch { url = `defaultuser/${extracted}` } } } // At this point url should be "owner/repo" url = `https://deepwiki.com/${url}` } normalizedInput.url = url } const parse = FetchRequest.safeParse(normalizedInput) if (!parse.success) { const err: z.infer<typeof ErrorEnvelope> = { status: 'error', code: 'VALIDATION', message: 'Request failed schema validation', details: parse.error.flatten(), } return err } const req = parse.data const root = new URL(req.url) if (req.maxDepth > 1) { const err: z.infer<typeof ErrorEnvelope> = { status: 'error', code: 'VALIDATION', message: 'maxDepth > 1 is not allowed', } return err } if (root.hostname !== 'deepwiki.com') { const err: z.infer<typeof ErrorEnvelope> = { status: 'error', code: 'DOMAIN_NOT_ALLOWED', message: 'Only deepwiki.com domains are allowed', } return err } // Progress emitter function emitProgress(e: any) { // Progress reporting is not supported in this context because McpServer does not have a sendEvent method. } const crawlResult = await crawl({ root, maxDepth: req.maxDepth, emit: emitProgress, verbose: req.verbose, }) // Convert each page const pages = await Promise.all( Object.entries(crawlResult.html).map(async ([path, html]) => ({ path, markdown: await htmlToMarkdown(html, req.mode), })), ) return { content: pages.map(page => ({ type: 'text', text: `# ${page.path}\n\n${page.markdown}`, })), } }, ) } - src/index.ts:30-30 (registration)Top-level invocation of deepwikiTool in the main server setup to register the tool with the MCP server instance.
deepwikiTool({ mcp } as McpToolContext)