Start free
Start free

In beta · 50 GB free, no card needed

Memory for everythingyour company knows.

Graphyte stores your documents, reads every page, maps how they connect, and answers questions with the source in hand. Four APIs, one bucket away.

Start freeRead the docs

Why Graphyte

Bigger models don't fix software that forgets. Graphyte reads what you store, relates it to everything else, and remembers the connections.

  • S3-style

    buckets, objects and keys you already know

  • Hybrid

    exact words and meaning in one query

  • Cited

    every answer points to its passage

Four APIs

Store files once. Search, answer and research over them from any app or agent.

01

Storage

Buckets, objects and keys, the way S3 does them. Put a file in and Graphyte reads it: PDFs page by page, images and screenshots, Markdown, HTML, JSON, CSV and plain text.

  • Overwrite by key, like S3
  • Indexed seconds after upload
  • Isolated per customer (tenants)
  • Up to 50 MB per file
store a file
curl -X PUT $GRAPHYTE/api/v1/buckets/ops/objects/charger-install.pdf \
  -H "Authorization: Bearer $GRAPHYTE_API_KEY" \
  -H "Content-Type: application/pdf" \
  --data-binary @charger-install.pdf

# → { "object": { "status": "pending", ... } }
02

Search

Exact words, meaning and the names of things in one query, then ranked by where each passage sits in the graph of everything you've stored. No index to configure.

  • Keyword + meaning + entities
  • Ranked across the whole graph
  • Scope to one bucket or all
  • Results in milliseconds
search
curl -X POST $GRAPHYTE/api/v1/search \
  -H "Authorization: Bearer $GRAPHYTE_API_KEY" \
  -d '{"query": "charger voltage"}'

# → each result says how it was found:
#   "via": ["keyword", "meaning", "graph"]
03

Answer

Ask in plain language, get a short answer where every claim points at the passage it came from. If your documents don't say, it tells you.

  • A citation on every claim
  • Says so when it doesn't know
  • Grounded in your files only
  • A few seconds per answer
answer
curl -X POST $GRAPHYTE/api/v1/answer \
  -H "Authorization: Bearer $GRAPHYTE_API_KEY" \
  -d '{"question": "What voltage does the depot charger need?"}'

# → segments of text, each with its citations
04

Research

For questions one search can't settle. Graphyte plans its own searches, follows names and numbers it finds into new ones, and writes a cited report.

  • Plans and runs its own searches
  • Flags gaps and conflicts
  • Shows every query it ran
  • Minutes, not days
research
curl -X POST $GRAPHYTE/api/v1/research \
  -H "Authorization: Bearer $GRAPHYTE_API_KEY" \
  -d '{"question": "Every charger incident this year, and why"}'

# → a cited report, plus "steps": the searches it ran

How it works

Search engines match passages one at a time. Graphyte also knows which passages belong together, across files.

  1. Store

    Put files in a bucket

    From the console, the API or your agent. Graphyte keeps the original and starts reading right away.

  2. Read

    Split into passages

    Every file becomes passages small enough to cite and big enough to make sense alone, each indexed by its words and its meaning.

  3. Relate

    Map what it mentions

    Graphyte writes down the people, products, sites and systems each passage names and how they relate, then links passages across files that share meaning or entities. The schema builds itself.

  4. Answer

    Rank by position, cite

    A question starts at the best matches and spreads along the links. Passages tied to many matches rise, so the answer can use what one search alone would miss.

Which voltage does the depot charger need, and is anything blocking the install?

400 V three-phase on its own 63 A breaker1. The install is blocked until the site survey is signed off3, which training notes say takes about a week2.

1 · manuals/charger-install.pdf, p. 142 · training/depot-onboarding.md3 · tickets/OPS-2291.md

One question, three files, found three different ways.

keywordcharger-install.pdfMatches the words in the question.
meaningdepot-onboarding.mdSays “power supply” and “sign-off”: different words, same meaning.
linkedOPS-2291.mdNever mentions the charger or its power. Found because it cites the manual's install section.

Built for

Anywhere people or agents need to ask what your documents say, and see where it says it.

  1. 01

    Company brain

    Policies, handbooks, decks and wikis in one place your team can ask. Answers link to the page, so people check the source instead of trusting a summary.

  2. 02

    Search inside your product

    Give your users search that understands what they mean across everything they've uploaded, without running Elasticsearch.

  3. 03

    Memory for agents

    Let your agents store what they read and look it up later with citations. One API key; works from any framework.

  4. 04

    Support that knows the manual

    Ground support replies in manuals, past tickets and release notes, and show the customer exactly where the answer came from.

Two calls to your first answer

Create a key in the console, upload a file, ask a question. Or hand the key to your agent: Graphyte speaks MCP, so it becomes your agent's memory in one command.

Read the API docs
quickstart.sh
# 1. Store a document
curl -X PUT $GRAPHYTE/api/v1/buckets/handbook/objects/leave-policy.md \
  -H "Authorization: Bearer $GRAPHYTE_API_KEY" \
  --data-binary @leave-policy.md

# 2. Ask about it
curl -X POST $GRAPHYTE/api/v1/answer \
  -H "Authorization: Bearer $GRAPHYTE_API_KEY" \
  -d '{"question": "How many days of parental leave do we offer?"}'
connect your agent (MCP)
claude mcp add --transport http graphyte https://graphyte.one/api/mcp \
  --header "Authorization: Bearer $GRAPHYTE_API_KEY"

Questions

How is this different from a vector database?

A vector database is one piece: you still have to store files, extract text, split it, embed it, and decide what to send the model. Graphyte does all of that from a file upload, mixes keyword and meaning search, and follows links between files, so related passages are found even when they share no words with the question.

Which file types does Graphyte read?

PDF, images (photos, screenshots, diagrams, scans: described and transcribed), Markdown, HTML, JSON, CSV, YAML and plain text, including source code. Audio and video are stored as-is and become searchable as readers for them arrive.

What is the ontology?

As Graphyte reads, it writes down the things your files mention and how they relate: who owns what, which site needs which part, which ticket blocks which install. Nobody defines the schema; types emerge from your documents. You can browse it in the console, and search uses it to connect passages that name the same thing.

Where does my data go?

Files are stored privately per workspace. To write an answer, only the passages that match the question are sent to the language model, for that one request. Your files themselves never leave your storage.

Can I use it from my agent?

Yes. Graphyte is an MCP server: add it to Claude Code, Claude Desktop or any MCP client with one URL and an API key, and your agent can search, answer, research and store notes it can recall later. Everything is also a plain HTTP call.

Can I build a multi-customer product on it?

Yes. Send an X-Graphyte-Tenant header with each request and every customer's files, search results, answers and links stay theirs, inside one workspace and one API key.

What happens when the answer isn't in my files?

Graphyte says so, and says what's missing, instead of guessing. Every claim it does make carries a citation you can open.

What does it cost?

Storage up to 50 GB is free while Graphyte is in beta, with no card needed. Paid plans for larger workspaces are on the way.

Give your software a memory.

Start with 50 GB free. Your first cited answer is about two minutes away.

Start freeRead the docs