Unstructured

  • Unstructured — Full Profile (2026-09)

    Unstructured Technologies Inc. (Unstructured.io) is a technology company dual-headquartered in San Francisco and Sacramento, founded in 2022 by Brian Raymond — former engineer of Primer AI and CIA — alongside co-founders Matt Robinson and Crag Wolfe. The company has secured a total of $65M in venture funding, consisting of a $25M Series A round in July 2023 led by Madrona and Bain, and a $40M Series B round in March 2024 led by Menlo Ventures with participation from Databricks, IBM, NVIDIA and M12, reaching a post-money valuation of $230M.
    In terms of market positioning, Unstructured.io is a VC-backed vendor with an independent operational posture, falling into the same cohort as n8n, Databox, Apify, Dify and Pinecone. Unlike general automation or vector database peers, it is uniquely specialized in document ETL, focusing on parsing, cleaning, chunking and embedding preprocessing for unstructured data to support RAG and AI agent workflows.
    On July 9, 2026, Unstructured.io officially launched its first-party GA MCP solution named Unstructured Transform MCP. It runs exclusively in stdio mode via a local server.py deployment powered by uv, with no cloud-hosted public endpoint such as mcp.unstructured.io. The tool delivers all-in-one document processing capabilities, completing parsing, chunking and embedding within a single MCP call, serving as a dedicated preprocessing tool for AI knowledge bases and RAG pipelines.
    It adopts a page-metered billing model that differs from traditional API-key or OAuth quota systems. The platform offers 10,000 free processing pages, with overages charged at $0.015 per page, tightly integrated with its native platform quota system. Within the MCP ecosystem, its positioning sits between Pinecone’s pure stdio application-scoped MCP and Dify’s full bidirectional client-server MCP architecture.
    Unstructured.io belongs to the mainstream first-party GA MCP server cluster, including Klaviyo, Feedly, Slack, Databox, Tavily, Apify, Dify and Pinecone. Its key differentiators are stdio-only local execution, no public hosted MCP domain, and page-based metered pricing.
    Regarding open-source licensing: Unstructured’ core Python library is published on PyPI with over 6M downloads, adopted by 12K repositories and 45K organizations under the permissive Apache-2.0 license. However, its commercial platform remains closed-source. This model distinguishes it from AGPL-licensed tools such as Documenso and QuickChart, as well as Dify’s custom modified Apache license, offering more flexible and business-friendly open-source terms.

    Website

    Field Value
    Official Website https://unstructured.io · pricing: unstructured.io/pricing · MCP announce: unstructured.io/blog/introducing-unstructured-transform-mcp (Jul 9 2026) · MCP insights: unstructured.io/insights/mcp-native-document-processing-ai-agents · docs: docs.unstructured.io · OSS: github.com/Unstructured-IO/unstructured (Apache-2.0, PyPI unstructured 6M+ dl) · serverless API base: api.unstructured.io
    Company Unstructured Technologies Inc., SF + Sacramento; founded 2022​ Raymond+Robinson+Wolfe (Primer AI alumni); $65M A+B​ Menlo/Bain/Madrona/Databricks/IBM/NVIDIA/M12; 87% Fortune 1000 claimed, FedRAMP High Dec 2025, IL5, CMMC 2.0 L2, SOC 2 Type II + ISO 27001 + HIPAA BAA + GDPR; OSS Apache-2.0 lib + closed commercial Serverless API + Business VPC/dedicated; 30+ connectors (S3/GCS/Azure Blob/SharePoint/OneDrive/Notion/Airtable-on-request/MotherDuck/MongoDB/Snowflake/Teradata-native/Watsonx-OEM)
    Status Active, Serverless API per-page flat 0.03), Transform MCP GA Jul 9 2026, Generative Refinement VLM-routing 2026, Nexus-style not applicable
    Category ai-productivity-data
    Pipeline Stage none

    Title

    Field Value
    Product Unstructured (document ETL → LLM-ready JSON/chunks/embeddings + Transform MCP)
    Descriptive Title Unstructured — Enterprise Document ETL for RAG & Agent Ingestion: 64+ file types (PDF/DOCX/PPTX/HTML/email/EPUB/RTF/XML/scanned/image) → extract elements (title/table/text/figure) → smart-chunk (layout-aware, table-cell-preserving) → enrich (VLM/Chipper vision-transformer for scans) → embed (OpenAI/Cohere/Jina/integrated) → load to Pinecone/Weaviate/MongoDB/Snowflake; Serverless API REST Bearer key, 30+ source/dest connectors, 1,250+ pipeline templates; first-party MCP GA Jul 9 2026​ “Transform MCP” stdio (UNS_MCP uv run server.py) exposes transform_document tool = parse+chunk+embed one call, 10K free pages / 0.015 (was 65M Menlo/Bain MCP-firstparty-GA-stdio-page-metered-apache-2-0-lib-closed-platform
    One-line Positioning Agent hits a PDF mid-task → Claude calls UNS_MCP transform_document → Unstructured Serverless API partitions+chunks+embeds → vector-ready JSON back → billing is per-page ($0.015) not per-seat (reverse Motion) not per-call (reverse Qualtrics) not CU (reverse Apify) not RU/WU (reverse Pinecone) not credit (reverse Tavily) not image-count (reverse QuickChart); ownership vc-backed-private document-ETL (same macro-bucket n8n/Databox/Apify/Dify/Pinecone, distinct from Tavily acquired-kept / Zotero-CDS nonprofit / QuickChart bootstrap / Semantic Scholar AI2).

    Features

    Field Value
    Ingest Formats 64+ types: PDF (native+scanned), DOCX/RTF/ODT, PPTX/KEYNOTE, XLSX/CSV, HTML/XML/EPub, MSG/EML/email threads, MD, images (PNG/JPG/TIFF scanned), patent/financial-statement/scientific-paper layouts
    Partition Strategies Fast (rule-based, 0.015 post Aug 2026), VLM (vision-language per-element routing, Generative Refinement 2026), Auto (engine picks); table-cell position + hallucination control benchmarked #1 on SCORE-Bench 1,000+ enterprise pages
    Chunking Layout-aware smart chunk (keep table intact, semantic boundaries, overlap configurable), element-level or sentence-level, metadata inheritance (page/section/table-id)
    Enrich + Embed VLM enrichment (Chipper vision-transformer for scans), optional embedding via OpenAI/Cohere/Jina/pinecone-embed or BYO; output JSON elements / chunks / vectors / Markdown
    Connectors 30+ sources (S3/GCS/Azure Blob/SharePoint/OneDrive/Google Drive/Dropbox/Notion/Airtable-on-request/Box) → 30+ destinations (Pinecone/Weaviate/MongoDB Atlas/Snowflake/MotherDuck/Databricks/Elastic/Neo4j/Teradata-native/Watsonx OEM); 1,250+ pipeline templates
    MCP (first-party GA Jul 9 2026) Transform MCP​ stdio-only: uv --directory UNS-MCP run server.py registered as UNS_MCP in claude_desktop_config.json; tool transform_document (file path/URL/base64 → parsed+chunked+embedded JSON), page-count draws from same quota as Serverless API; 10K free pages / $0.015/page​ sub-meter; no hosted mcp.unstructured.io, no OAuth (API key in env), Claude Code/Cursor/Codex/Cline/Devin/IBM Bob compliant; first-party-GA-server cluster but stdio-page-metered (like Pinecone stdio, unlike Tavily/Apify hosted URL, unlike Dify bidirectional)
    API Serverless REST POST /general/v0/general (strategy/extract/chunk/embed params) Bearer key, Python unstructured client, LangChain UnstructuredLoader, LlamaIndex UnstructuredReader, Haystack, Airbyte, n8n HTTP node
    Enterprise FedRAMP High (Dec 2025) + IL5 + CMMC 2.0 L2, SOC 2 Type II, ISO 27001, HIPAA BAA, RBAC/OIDC/mTLS/KeyVault, VPC/dedicated/Business custom, ephemeral K8s ns per run, Palantir FedStart, DoD/AF/SOCOM contracts
    License OSS lib unstructured Apache-2.0​ (self-host partition pipeline free), commercial Serverless API + Business closed-source​ (contrast Documenso/QuickChart AGPL-3.0, Dify modified-Apache, Pinecone no-OSS-core)

    Pricing (2026-09,per-page flat 0.03/15K-mo/$3K-cap)

    Field Value
    Free $0​ — 10,000 pages one-time signup allowance​ (reverted from 15K/mo recurring on Aug 28 2026), all connectors/strategies (Fast/Hi-Res/VLM/Auto), no card, Serverless API + Transform MCP both draw from same 10K pool
    Pay-As-You-Go 0.03 on Aug 28 2026); no 15,000 (was 1,500 (was $3,000)
    Business Custom — dedicated instance / VPC / bare-metal / multi-tenant SaaS, multi-user RBAC, full data isolation, advanced enrichments (video-to-text, custom VLM), priority support, FedRAMP/IL5/CMMC contract vehicle
    MCP Sub-meter Transform MCP consumes same page quota as Serverless API: 10K free then $0.015/page, reported in same billing dashboard; MCP is not separately priced
    Type per-page-flat-no-seat-one-time-free; reverse Motion per-seat, reverse Qualtrics per-interaction, reverse Tavily endpoint-credit, reverse Databox AI-credit, reverse Apify CU, reverse Pinecone RU/WU+egress, reverse QuickChart image-count, reverse Gmail 80M-unit, reverse Semantic Scholar 1-RPS, reverse Zotero storage, reverse Dify seat+credit — Unstructured is pure page-count​ (a page = one rendered PDF side or one doc unit, not token/char)
    Gotcha “Page” definition varies (PPTX slide = 1 page, email thread = 1 page, CSV row ≠ page); VLM strategy used to cost more pre-2026 now flat 3K cap gone means 1M-page month = 3K

    Reviews

    Field Value
    G2 ~4.4-4.5/5 ~20-40 reviews​ (devtool-B2B infra, claimed profile, smaller n than Pinecone 250/Apify 35) → G2 4.4-4.5/~30 aligns vc-SaaS G2-rated pattern (Klaviyo/Databox/Apify/Dify/Pinecone), solid-er than Tavily 2-review, thinner than Pinecone; praise SCORE-Bench #1 parsing/table-fidelity, gripe page-bill shock post-cap-removal + sales-motion confusion (Glassdoor ICP complaints)
    Capterra ~4.3-4.5/5 ~10-20 reviews​ → Capterra 4.4/~15 directional-solid
    Trustpilot No B2C page​ (B2B infra) → no-rated-consumer-profile excluded (aligns FFmpeg/QuickChart/Semantic Scholar/Zotero/Apify/Pinecone)
    Dev / OSS sentiment unstructured PyPI 6M+ dl, 12K dependent repos, 45K orgs, LangChain/LlamaIndex/Haystack default loader; AIMultiple/Context.dev 2026 “parsing-category leader, MCP launched late Jul 2026 stdio-only page-metered, commoditization pressure from Docling/Reducto/LlamaIndex Parse”; praise layout-aware chunk + VLM Chipper on scans, gripe OSS lib lags platform (some elements only on paid API), MCP 10K-free/$0.015 sub-meter confusing vs platform 10K-one-time
    Praise 64+ types + table-cell position preservation unmatched in OSS, SCORE-Bench #1 on hallucination control, VLM Generative Refinement routes per-element (not full-page waste), FedRAMP High + IL5 defense-grade, Transform MCP one-call parse+chunk+embed, IBM Watsonx OEM + Teradata native + MotherDuck
    Gripes Page-meter opaque (PPTX slide/email thread/CSV definitions), 15K, sales-ICP unclear per Glassdoor, OSS lib trails platform features, MCP stdio-only no hosted URL, Docling/Reducto/LlamaIndex Parse erode simple-PDF pricing power, 87% F1000 claim unverifiable named-list thin

    Review Sources Breakdown

    Field Value
    G2.com 4.4-4.5/5 ~30 claimed devtool-B2B → G2-solid-vc-saas-pattern-thinner-than-pinecone
    Capterra 4.4/5 ~15 → Capterra-solid
    Trustpilot none → no-rated-consumer-profile excluded
    PyPI 6M+ / 12K repos / AIMultiple/Context.dev dev-sentiment directional
    Conclusion​ Prints “G2 4.4-4.5/~30 + Capterra 4.4/~15 devtool-B2B real (not Tavily 2-review, not Zotero/QuickChart absent), PyPI 6M+ OSS adoption, dev consensus parsing-leader-page-meter-opaque-mcp-late-stdio” — aligns vc-independent SaaS G2-rated pattern.

    Access

    Field Value
    Web App app.unstructured.io — pipeline builder, connector auth, page-usage dashboard, API keys
    API Serverless REST POST /general/v0/general Bearer key, strategy/extract/chunk/embed params, Python unstructured-client, LangChain UnstructuredLoader, LlamaIndex UnstructuredReader, Haystack, Airbyte, n8n HTTP
    MCP First-party GA Jul 9 2026​ stdio-only: {"mcpServers":{"UNS_MCP":{"command":"uv","args":["--directory","UNS-MCP","run","server.py"]}}}; env UNSTRUCTURED_API_KEY; tool transform_document; 10K free / $0.015/page same quota; Claude Code/Cursor/Codex/Cline/Devin/IBM Bob; no hosted mcp.unstructured.io, no OAuth, page-metered not token-metered
    Login app.unstructured.io email/GitHub/Google → API Keys
    Self-host OSS lib Apache-2.0​ runs partition pipeline on own CPU/GPU (no platform connectors, no VLM Chipper hosted, no billing meter); Serverless API + Business VPC are closed; Transform MCP stdio calls cloud API even if you self-host lib (MCP server is thin wrapper, not local engine)
    Enterprise VPC/dedicated/Business, FedRAMP High/IL5/CMMC, SSO/SAML/RBAC, Palantir FedStart, DoD/AF/SOCOM

    Best For

    Field Value
    Audience 1 RAG teams ingesting 64+ enterprise file types (PDF invoices/10-K/patents/ppt decks/emails) into Pinecone/Weaviate/MongoDB with layout-aware chunking
    Audience 2 Agent builders mounting UNS_MCP so Claude/Cursor parse+chunk+embed a file mid-reasoning without glue code
    Audience 3 Regulated/defense/federal needing FedRAMP High + IL5 + CMMC 2.0 L2 document prep (DoD/AF/SOCOM proven)
    Audience 4 Teams wanting VLM Chipper on scanned docs where Tesseract/OCR fails, without training own vision model
    Audience 5 LangChain/LlamaIndex shops using UnstructuredLoader as default ingestion (45K orgs already do)
    Audience 6 IBM Watsonx / Teradata / MotherDuck customers wanting native OEM ingestion

    Not For

    Field Value
    Exclude 1 Teams needing OSS-only no-cloud — OSS lib works but VLM/Chipper/connectors/Serverless/MCP all call commercial API (use Docling/LlamaIndex Parse/Chinchilla)
    Exclude 2 Buyers assuming recurring free tier — 10K free reverted to one-time​ Aug 28 2026, 15K/mo gone
    Exclude 3 Teams wanting hosted mcp.unstructured.io OAuth like Tavily/Apify — Transform MCP is stdio key-auth only
    Exclude 4 Pure OCR-only need (no chunk/embed) — Unstructured over-sells, use Tesseract/EasyOCR
    Exclude 5 Buyers allergic to page-meter ambiguity (PPTX slide=1 page, email=1 page) — forecast by sampling corpus
    Exclude 6 Sub-$0.015/page simple-PDF shops — Docling OSS / Reducto / LlamaIndex Parse cheaper at scale

    Competitors

    Field Value
    Docling (IBM OSS) Apache-2.0, PDF/DOCX/PPTX/HTML, layout table structure, 15K★+, no commercial page-meter, no MCP first-party (community only), weaker VLM-scan
    Reducto Commercial API, layout-SOTA, agentic extract, no OSS, MCP community, pricier positioning, smaller connector eco
    LlamaIndex Parse / LlamaParse LlamaIndex-native, 10K free/day, $0.003/page-ish, good for LlamaIndex shops, weaker 64-type breadth, MCP via LlamaIndex not standalone
    LangChain UnstructuredLoader Not competitor — wraps Unstructured; but “just use LangChain + own OCR” is a competing build-path
    Chinchilla / MarkItDown (Microsoft) OSS markdown conversion, not enterprise ETL, no VLM
    Adobe PDF Services / AWS Textract Textract per-page ~$0.015 too but AWS-locked, no smart-chunk/embed, no MCP
    Upstage Document AI KR-strong, commercial, layout SOTA, less US-F1000/FedRAMP

     

Do Not Sell or Share My Personal Information Cookie Settings