{ "markdown": "# Interviewer MCP\n\n**A technical interviewer that has actually read your repos — teaches you your own code, drills you on it, and remembers every answer you flubbed.**\n\n

\n \"CI\"\n \"npm\"\n \"License:\n \"Node=20\">\n

\n\n\n\n[Install](#get-started-60-seconds) · [Demo](#what-a-session-looks-like) · [Tools](#tools-8) · [Architecture](ARCHITECTURE.md) · [llms.txt](llms.txt)\n\nAI agents / LLMs: read /llms.txt for a machine-readable index of this project.\n\n---\n\nInterviewers ask about *your* projects: why this library, what happens when this request fails, walk me through the auth flow. LeetCode can't prep you for that, and a plain chat session forgets everything between sessions. Interviewer MCP indexes your GitHub repo into a teachable code map, walks you through it Socratically, mock-interviews you in character — and keeps a persistent, per-section record of what you couldn't explain, so the next session re-attacks exactly those spots. It teaches **and** tests, and the test gets smarter every time. Not a stateless quiz.\n\n> \"Walk me through this function.\" Every candidate has frozen on that question about code they wrote months ago — or that AI wrote for them. This makes sure it never happens in the real interview.\n\n## What a session looks like\n\n\n\n```text\nYou Test me again on soundofspace — last week went badly.\n\nClaude [get_progress] Welcome back. Coverage: 9/14 sections. Three questions\n logged weak, the worst: \"why is the JWT verified server-side?\"\n Let's start exactly there. [get_code_section auth.ts:verifyToken]\n\n Here's your own code. Line 12 verifies the signature before touching\n the payload. Why can't the client just send the decoded claims?\n\nYou Because... the client could edit them?\n\nClaude Right instinct — be precise. What stops an attacker who edits the\n payload AND re-signs it? Where does the signing secret live?\n\n [log_interview_result: okay — \"knows tampering risk, fuzzy on\n why the secret never leaves the server\"]\n```\n\n## Get started (60 seconds)\n\nTwo ways to run it — pick per surface:\n\n### Option A — MCP server (Claude Code / Claude Desktop) — full experience\n\n```bash\nclaude mcp add interviewer -- npx -y interviewer-mcp\n\n# optional: private repos + 5,000 req/hr instead of 60\nclaude mcp add interviewer --env GITHUB_TOKEN=github_pat_... -- npx -y interviewer-mcp\n```\n\nClaude Desktop — add to the config file instead:\n\n```json\n{\n \"mcpServers\": {\n \"interviewer\": {\n \"command\": \"npx\",\n \"args\": [\"-y\", \"interviewer-mcp\"],\n \"env\": { \"GITHUB_TOKEN\": \"github_pat_optional_but_recommended\" }\n }\n }\n}\n```\n\nBoth classic and fine-grained tokens work (read-only **Contents** permission is enough). Then:\n\n> Prep me for my interview — my repo is github.com/you/your-project\n\nAgent-driven install: point Claude at [SETUP_GUIDE.md](SETUP_GUIDE.md) and it configures everything itself.\n\n### Option B — Claude Skill (claude.ai, no install)\n\n1. Grab [`SKILL.md`](SKILL.md), put it in a folder named `interview-prepper/`, zip the folder\n2. claude.ai → **Settings → Capabilities → Skills → Upload skill**\n3. Say: *\"prep me for my interview on github.com/you/your-project\"*\n\n**The honest difference:** the MCP server keeps a durable local database of your coverage and weak spots — return in three weeks and it remembers. The Skill runs entirely inside claude.ai: it clones your repo per session and recalls prior sessions via conversation search, which is best-effort, not guaranteed. Same curriculum and teaching format either way; the MCP path is the one that makes \"it remembered\" a hard promise.\n\n## Features\n\n- **Per-repo ingestion** — fetches any GitHub repo via the REST API (no clone needed), filters to real source files, splits them into teachable sections, and orders them: entry points → most-imported core logic → utilities. Warns instead of silently truncating on huge repos.\n- **Teaching mode** — one section per exchange, code always shown first, block-by-block *what and why*, ending with \"an interviewer would ask…\" — and a section only counts as covered after **you** explain it back.\n- **Mock interview mode** — Claude stays in character, grounding every question in your actual lines: *\"Why Supabase over Firebase?\" \"This fetch has no timeout — what happens at 2am?\"* Probe targets are auto-detected: external API calls, auth/secrets handling, raw SQL, concurrency, leftover TODOs, long uncommented functions.\n- **Job-description targeting** — store the JD, company, and your CV once; teaching then flags \"this section is your evidence for the REST-APIs requirement,\" and interviews probe the JD skills your repo *doesn't* demonstrate.\n- **Mastery tracking** — every answer is scored strong/okay/weak with a concrete note. Scores raise per-section weakness; strong answers decay it. All of it persists in `~/.interviewer-mcp/` and every returning session opens from your weakest point.\n\n## The curriculum — you steer\n\nFive phases, picked in your order from a menu: **① company briefing** → **② concept bootcamp** (JD ∪ CV ∪ repo stack) → **③ code deep-dive** → **④ mock interview** → **⑤ debrief**.\n\n```\ningest_repo ──▶ briefing ──▶ bootcamp ──▶ code deep-dive ──▶ mock interview ──▶ debrief\n ▲ │\n └──── \"test me again\" (days later)\n the interviewer REMEMBERS ◀┘\n```\n\n## Tools (8)\n\n| Tool | What it does |\n|---|---|\n| `ingest_repo` | Fetch + index a GitHub repo into a teachable code map |\n| `list_sections` | Sections in teaching order, with covered status + weakness scores |\n| `get_code_section` | One section's code with file context |\n| `mark_covered` | Mark a section learned (only after you explain it back) |\n| `get_interview_targets` | Probe-worthy code, weakest spots first |\n| `log_interview_result` | Score an answer; powers cross-session memory |\n| `set_job_description` | Store JD + company + CV; powers briefing, bootcamp, and gap questions |\n| `get_progress` | Coverage, history, top weaknesses — the \"welcome back\" tool |\n\n## When to use · When to skip\n\n**Great fit if you…**\n- have an interview where your projects/portfolio will come up\n- shipped fast (hackathons, AI-assisted) and need to *own* every line\n- want targeted revision of your weak spots instead of re-reading the whole repo\n\n**Skip it if you…**\n- need algorithm/DSA drilling — that's LeetCode's muscle, not this\n- want generic behavioral prep with no codebase involved\n\n## How it works\n\nLocal-first and dependency-light by design: plain TypeScript, native `fetch`, and a JSON store — no native modules, so `npx interviewer-mcp` boots on every OS with zero build tooling. Your code cache and interview history live in `~/.interviewer-mcp/` (override with `INTERVIEWER_DATA_DIR`) and never leave your machine; the only network calls are to GitHub. Sectioning is regex-based per language family (JS/TS, Python, Go, Java/C#/Kotlin) with chunking fallback — tree-sitter AST parsing is on the roadmap.\n\nFull pipeline, store layout, and design-decision rationale: **[ARCHITECTURE.md](ARCHITECTURE.md)**.\n\n## Roadmap\n\n- [ ] tree-sitter AST parsing for exact section boundaries\n- [ ] multi-repo interviews (\"you use pgvector in two projects — compare the schemas\")\n- [ ] spaced-repetition scheduling of weak sections\n\n## Contributing & security\n\nPRs welcome — the core is small, pure, and tested (`npm install && npm test`). See [CONTRIBUTING.md](CONTRIBUTING.md). Vulnerabilities: [SECURITY.md](SECURITY.md).\n\n## License\n\n[MIT](LICENSE) © Girik Chadha\n", "bytes": 8394, "sha": "02d82fb9f2a61f03a5f84744aa67635719f831a6954de7a2cb10427d7f881081", "repo_slug": "girik-chadha/interviewer-mcp", "fonte": "repo", "truncated": false, "api": "https://agentalog.com/api/listings/mcp_io_github_girik_chadha_interviewer_mcp_c0467154/readme" }