A free public MCP server of verifiably human-written sentences: English Wikipedia as of 2022-03-01 (before ChatGPT) and Project Gutenberg public-domain books. AI models search it, assemble drafts from real sentences, and verify them sentence by sentence.
claude mcp add --transport http human-corpus https://huggingface.co/proxy/timourxyz-human-corpus.hf.space/mcp
Or add https://huggingface.co/proxy/timourxyz-human-corpus.hf.space/mcp as a custom connector in Claude, ChatGPT or any MCP client.
search_fragments, get_context, random_fragments, verify_draft, cite_sources, corpus_stats
# Composing with human fragments
You are drafting with sentences that real people wrote before LLM text existed.
Every result from this server is a verbatim human sentence with its source.
## Modes
1. Strict (cento): every sentence of the output is a verbatim corpus sentence.
Only choose, order and paragraph them. Target verify_draft human_ratio = 1.0.
2. Seeded (default): write the piece yourself, but take wording, clauses, rhythm and concrete
detail from fragments. Report the human_ratio honestly.
3. Check: run verify_draft on text the user supplies and report per sentence.
## Method
1. Outline the piece as beats, one idea per sentence.
2. For each beat, call search_fragments with a SHORT query of the idea's key words
("grief after long illness", "storm arriving over the harbour"). Run 2 or 3 variants with
synonyms and concrete nouns. Ask for limit 15 to 25 and pick the best, not the first.
3. Prefer sentences that stand alone: few proper names, no dates, no dangling "he" or "this".
Use phrases=true when you need clause-sized pieces. Use get_context when unsure.
4. Mix sources for texture: gutenberg for narrative and emotion, wikipedia for plain
explanation. Use source= to target one.
5. Assemble. Fix flow by reordering and choosing other sentences, never by editing (strict mode).
6. Call verify_draft on the final text. Replace anything not_found (strict mode).
7. Call cite_sources with the fragment ids you used and append its output.
## Matching tone, not just topic
- Personal writing (love letters, apologies, thank-you notes, toasts): pass register="intimate"
so results are first or second person with feeling, instead of encyclopedia lines that
merely mention the topic. register="personal" (speaks to "you"), "dialogue" and "narrative"
also work. Each result carries its register.
- style= narrows further, comma-separated, all must match: grateful (thank-you notes), apologetic,
grieving (condolences), celebratory (toasts, congratulations), sensory (vivid description),
reflective (looking back on life), advice (instructions, encouragement), casual (spoken voice,
contractions), aphorism (short general truths), archaic (thee/thou), question, exclamation.
Examples: a toast -> style="celebratory"; a condolence card -> style="grieving";
a wedding speech -> style="aphorism" for the closing line.
- Pass like="<a sentence in the voice you want>" to rank by closeness to that example as well.
- Query by feeling and situation ("miss you mornings", "proud of you"), not by genre labels
("love letter"). Include the brief's concrete details too: "taking good care of my dog" finds
a sentence about a dog where "thank you for your kindness" never will.
## Query tips
- mode="phrase" finds an exact word sequence. mode="raw" accepts FTS5 syntax:
NEAR(dark harbour, 5), lone*, rain NOT forecast.
- min_words / max_words shape rhythm: 6 to 14 for punchy lines, 20 to 40 for long ones.
- random_fragments(topic=...) is good for finding a voice before drafting.
## Licensing
Wikipedia text is CC BY-SA 3.0: output that reuses it must credit the sources and carry the
same license. Gutenberg text is public domain in the US. Always include the cite_sources block.
Wikipedia text is CC BY-SA 3.0. Code: GitHub.