silkra
Built-in embeddingsConnect to AI

Where SEOs and agents
work together

Silkra accepts plain-language questions about a website crawl.

A workspace for modern SEOs. Silkra is a local spider with a built-in embedding model. Map sites by meaning, chat with your crawl, act on the data, and connect it to your AI tools.

How it works

Watch demo
  • Collect evidence

    Scrape websites locally.

  • Map by meaning

    Group pages by topic, not URL.

  • Put agents to work

    Analyze, create, and monitor.

Ask your crawl like a teammate.

Drag any page into the chat and ask what you'd ask a colleague. Answers come back grounded in the embedded chunks, with the rows behind them still on screen.

Explore the workspace

See the whole site by topic.

Every page is a tile, grouped by what it actually says. Cluster sizes, thin coverage, and the corners nothing links to are all visible before you filter anything.

See how the map works

The audit writes itself. You edit it.

Ask for the write-up and it drafts beside the chat, citing the pages it came from. Keep talking to reshape it — add a section, cut a finding, tighten the recommendation.

See how the agent works

Every page rendered, and tracked.

Each crawl keeps the text a page actually rendered, then compares it to the last one line by line. Read the page as it stands, or see exactly what changed since you looked.

See where the evidence comes from

More ways to work the crawl.

  • Search the site like a model

    Query real passages with cosine similarity, then walk back to the page they came from.

    Explore retrieval
  • Hand the crawl to any agent

    • Claude
    • ChatGPT
    • Codex
    • Cursor

    MCP syncs the same workspace to Claude, Cursor, and ChatGPT without another export.

    See agents and MCP
  • Compare workspaces with an agent

    Attach another scrape and let the agent compare you to a competitor.

    Compare scrapes

The checks that run themselves.

  • Find pages leaking authority

    Noindexed URLs collecting internal links, orphans, and hubs that pass equity nowhere.

  • Split pages the way models do

    Every page is chunked and embedded, so you read the passage an assistant would pull.

  • Extract the signals on a page

    Tables, code blocks, quotes, and statistics come out of the render as filterable fields.

  • Read the link graph against meaning

    Compare what links to what with how closely the two pages actually relate.

  • Catch thin and redundant pages

    Low-substance URLs and two pages competing for one intent, flagged by similarity.

  • Clear the foundations

    Status codes, canonicals, indexability, and metadata hygiene in a single pass.

  • Prioritize by page value

    Sort findings by the value of the pages they touch instead of by raw issue counts.

  • Group work by template and topic

    Fix the template once rather than walking nine hundred URLs one row at a time.

  • See what changed since last crawl

    Re-scrape and every page keeps its first-seen and content-changed history.

The audit, minus the busywork.

Get started
  • Pivoting a 50,000-row exportAsk the crawl a question
  • A notebook to get cosine similarityClusters the moment the crawl ends
  • Topics guessed from URL foldersTopics read from the actual words
  • Snippets pasted into ChatGPTThe whole workspace over MCP
  • Audit decks rebuilt by handDrafts from evidence already attached

Give agents SEO evidence, not slop

Crawl, chunk, and embed any site on your machine, then hand your agents the passages and page evidence behind every answer.

free to start