TeachClaw

A teaching assistant your students can chat with on Telegram. It draws on your slide deck, the readings you assign, the question bank you've written, and (soon) whatever's on screen at that moment — and cites the source for every claim it makes.

Think of it as the patient TA who does the office hours you can't: each student gets their own private thread, asks the dumb question they were too embarrassed to raise in class, and gets back an answer grounded in your materials, not a guess from the open internet.

What it's good for

• Pre-class warm-up. Students arrive having already poked at the reading; the bot answers definitions, framework questions, and "wait, what does that acronym mean" so class time is for harder stuff.
• Live Q&A backchannel. During the lecture, students who don't want to raise their hand drop questions into the bot. The bot can see your slides and answers in context; you see the volume and themes during a break.
• After-class deepening. "Walk me through the two-by-two on slide 14 again" or "give me the McKinsey-flavored read of the case." The bot can switch personas (Skeptical CFO, VC Partner, McKinsey Analyst, Professor) so the same material can be probed from different angles.
• Defensible answers. Every substantive claim links to the source — your slide, the assigned PDF, the YouTube talk at the exact timestamp. If the bot can't find a source, students learn to push back.

Cost

Designed to run on inexpensive open-weight models — Google's Gemma 4 26B A4B ^$0.06/$0.33, Gemma 4 31B ^$0.13/$0.38, or Qwen 3.5 Flash ^$0.07/$0.26. Forty students should cost under $2/hr at API rates. Aggressive prompt caching keeps it there.

Sample knowledgebases are on Github: AI in Business, Agentic AI.

This is a branch of BenchClaw, which is itself a fork of nanobot, though very little of the original code remains.

Running

uv run teachclaw

Config file: config.yaml, created automatically on first run. By default it's tuned for google/gemma-4-e4b-it; small models can have quirks.

Connecting a knowledgebase

The course library lives in a separate repo and is exposed to the bot as a Model Context Protocol (MCP) server. The bot can call into it the same way it calls any other tool, but the corpus is curated by you — slides, readings, videos, question bank — and the bot is restricted to it. No open-web browsing unless you wire that in deliberately.

Wire it up by pointing mcp_servers at the knowledgebase's launch command:

mcp_servers:
  - name: kb
    transport: stdio
    command: sh
    args:
      - "-c"
      - "cd /path/to/lecture-knowledge; uv run knowledge-mcp"

Or, if you'd rather run the knowledgebase as a long-lived service:

mcp_servers:
  - name: kb
    transport: http
    url: http://127.0.0.1:8765/mcp

The reference knowledgebase (lecture-knowledge) includes:

• The lecturer's slide decks (workshop backbone).
• Macro AI context — AI Index, Epoch notable models, OWID AI page — for benchmarks and capability claims.
• Vendor pricing for model comparisons.
• AI regulation for governance and policy questions.
• Selected consulting reports, essays, and canonical YouTube talks.
• Memes for the light-touch teaching layer.

To swap in your own corpus, fork the knowledgebase repo, replace the fetch/process manifests with your sources, and re-run the build. The bot doesn't care what's in there as long as the MCP surface stays the same.

Citations

Every substantive answer links back to the source it came from. The bot emits <citation id="…"> markers around any claim it sourced from the knowledgebase; the channel layer rewrites those into clickable references the student can tap.

A citation looks like:

• [NVDA 10-K FY26 / Item 7] — the MD&A section of that filing
• [McKinsey State of AI 2025, p.14] — page 14 of the PDF
• [Karpathy "Intro to LLMs", 12:45] — the video at 12:45
• [Sequoia: AI Ascent 2026 Keynote] — the linked talk

Tapping the reference opens the original document at the right page, section, or timestamp. Sourcing is enforced server-side: if the bot makes a strong claim without a citation, the channel pushes back and asks it to cite or retract. Students who want to dig in can react to any reply with ❤ to see the full source list, or with 🔍 to see the tool-call trace that produced the answer.

Pending features

• Screen scraping. Capturing the lecturer's screen (the current slide, a live spreadsheet, a code editor) so the bot can answer "what's on screen right now?" questions in real time.
• Interactive Q&A. Bot-driven mini-quizzes — the bot picks a concept from the question bank, asks the student, and walks them through their answer with hints and follow-ups instead of just grading.