Extend
Document processing infrastructure for AI agents: five composable APIs for parsing, extraction, splitting, classification and form editing, wrapped in versioned workflows with conditional routing, a confidence-threshold human review gate, and an evaluation framework with labeled evaluation sets and a publish gate.
Extend is document processing infrastructure built for the documents that break older tools, such as complex tables, merged cells, signatures, handwriting and degraded scans, and rebuilt on vision language models instead of legacy OCR. It was founded in 2023 by Kushal Byatnal and Eli Badgio, who built data quality and validation pipelines at Flatiron Health.
It ships as five APIs that work alone or together: parse a document into LLM-ready markdown, extract structured fields, split a bundle into its constituent documents, classify what each one is, and edit or fill forms. An agentic pipeline routes pages through vision models, and a Review Agent re-checks low-confidence extractions before they are accepted.
Around those sit the parts that make the output trustworthy in production. Workflows compose the APIs into pipelines with routers, conditional steps, formulas and validation, including a step that checks an extracted value against the customer's own system of record mid-run, with versioning and durable async and batch execution.
An evaluation framework lets a team build labeled evaluation sets, score a configuration at field and document level, catch regressions and only then publish it. Composer, a background agent, generates an extraction schema from a handful of sample documents and keeps refining it against those scores. Extractions below a confidence threshold route to a human review interface, which can be embedded in the customer's own product.
Developers reach it through a REST API, SDKs, a CLI, an MCP server and webhooks. Processing runs in one of three named regions, us1, us2 or eu1, the last in Frankfurt. Extend is HIPAA compliant with business associate agreements available, GDPR compliant, and maintains controls aligned with SOC 2 Type II with reports in its trust center. Billing is credit-metered on document volume. Customers include Brex, Square, Checkr and Flatiron Health.
Vendor details
Canonical URL
https://extend.ai
Category
Agent infrastructure
Subcategory
Document processing infrastructure — agentic parsing, extraction and evaluation
Company status
independent
Use cases & customers
Deployment options
Integrations
Extend is infrastructure a customer's engineers wire in rather than a connector platform, and it publishes no catalog of prebuilt business connectors. Outbound reach is per-mechanism and documented: an External Data Validation workflow step calls the customer's own system mid-run to check an extracted value before the pipeline continues; webhooks push run outcomes outward with a documented event catalog; and a GitHub App integration is documented under Workflows. Inbound, the platform is reached by REST API with a published reference, SDKs, a CLI and an MCP server, with embeddable UI Components putting Extend's review interface inside a customer's own product. The counterparties documented in the product are file types rather than applications — the General section documents Supported File Types where a connector-led platform would document connectors.
In practice
A fintech team needs to pull fields from thousands of bank statements and loan applications with near perfect accuracy. Extend classifies each document, extracts the data, and flags only the low confidence cases for human review.
Your operations team is drowning in handwritten forms and degraded scans that legacy optical character recognition mangles. Extend uses vision language models to read and correct them, returning clean structured data ready for your systems.
An engineering team wants to stop maintaining brittle extraction scripts. They upload sample documents to Extend Composer, which generates a tuned schema, runs evals, catches regressions, and improves accuracy in the background.
Sources & related URLs
Agentic Index coverage score
9.0 / 14 capabilities · 64%
| Integrations & Tool Calling | Partial |
|---|---|
|
The External Data Validation Step calls a customer's own system mid run to check an extracted value, an outbound call inside a running process and the clearest form of tool use here. Webhooks push run outcomes into the customer's systems, with configuration options and an event catalog. A GitHub App integration is available for Workflows. UI Components let Extend's review interface be embedded in a customer's own product. The platform is reachable by REST API, SDKs and a CLI, so a customer wires it wherever needed. There is no catalog of prebuilt business connectors, such as Salesforce, Workday, NetSuite or SharePoint. Extend is document processing infrastructure a customer's engineers integrate, not a platform pre wired to their systems, and the named counterparties are file types, not applications. The agent acts on documents and hands structured data back to the caller. The Edit API fills forms and generates documents, so it too acts on a document, not a system. Sourcedocs.extend.ai Workflows, Webhooks and General navigationread 2026-09-12 |
|
| Workflow Orchestration | Full |
|
Workflow step types include Parse Step, Validation Step, Router Step, Conditional Steps, External Data Validation Step and Conditional Extraction Step, which make this orchestration, not a pipeline. A router and conditional steps are branching primitives, so the run takes different paths depending on what the previous step found. Formulas add computation between steps. External Data Validation reaches outside mid run. A workflow pauses to check an extracted value against the customer's own system of record before continuing, holding state across a call the platform does not control. Durability and versioning are built in. Workflow Versioning and per run inspection through Reviewing Workflow Run are available, and Async Processing and Batch Processing are separate, so a run survives beyond a request, at volume. Multifile Extraction in Workflows handles one logical document arriving as several files. The five APIs compose instead of stacking. Parse, Extract, Split, Classify and Edit work alone or together, and Extend contrasts this with rivals where multi-step workflows require custom engineering because capabilities are bundled into a single endpoint. The participants are distinct. They are the pipeline, the Review Agent correcting low confidence extractions, the human reviewer at the confidence gate, and the customer's external system answering a validation call. Sourcedocs.extend.ai Workflows navigation with extend.ai/resources document processing APIs writingread 2026-09-12 |
|
| Knowledge Grounding & RAG | Not documented |
|
The product sits upstream of retrieval. Extend calls itself document processing infrastructure for AI agents, and its job is to convert unstructured documents into structured, LLM-ready markdown. It produces the grounding material that someone else's retrieval system consumes, and manufacturing the corpus is not the same as being grounded in one. Extend's work is per document. The five capabilities (Parsing, Extraction, Splitting, Classification, Editing) operate on the file in front of them. There is no index, semantic search surface, vector store, accumulated corpus a customer queries, or citation back to source. Multifile Extraction handles one logical document arriving as several files, a batching concern, not a corpus. What persists is configuration, not knowledge. Extraction schemas, workflow versions, evaluation sets and processors are versioned, but they describe how to read a document, not what the organization knows. Evaluation sets are labeled ground truth for scoring, and Memory sits under Optimization. None of these retrieval pieces appears among Extend's named capabilities. Sourcedocs.extend.ai Capabilities navigation with extend.ai homepageread 2026-09-12 |
|
| Human Oversight & Guardrails | Full |
|
The review gate spans three surfaces. Confidence Scores are the signal the gate fires on. A Review Agent uses vision language models to review and correct low confidence extractions. Reviewing Workflow Run is the human facing surface where a run is inspected before its output is accepted. Extend also offers a human in the loop review UI alongside confidence scoring and LLM-as-a-judge scoring. The gate is on the agent's own output. An extraction below the confidence threshold does not flow downstream. It is routed to a person who inspects and corrects it. For a product whose value is that extracted data is trustworthy enough to write into a finance or healthcare system, the hold before acceptance is the product. Two validation steps, Validation Step and External Data Validation Step, are workflow primitives that extend the gate to anywhere in a pipeline. The second validates an extraction against the customer's own system of record before accepting it, an automated check that can block a run, placed where the customer chooses. Confidence threshold routing fires on a measured signal, not a rule someone remembered to set. Sourcedocs.extend.ai extraction and workflows navigation, with extend.ai/resources document processing APIs writingread 2026-09-12 |
|
| Security, Identity & Governance | Full |
|
Extend's SOC 2 claim is hedged. Extend maintains controls aligned with SOC 2 Type II requirements, with reports and security documentation in its trust center. Aligned with is weaker than certified. The other two claims are unhedged, and one is contractual. Extend is HIPAA compliant and can execute Business Associate Agreements (BAAs) across all three deployments. A signable BAA is a legal instrument accepting liability as a business associate, and HIPAA has no certificate to hold. Extend also states flatly that it is GDPR compliant, with a complaint procedure that sets acknowledgment within 3 business days and a response within 30. There are user roles and permissions, API authentication, and data handling rules covering storage, retention and deletion. Per region deployment isolation adds a data boundary control. The SOC 2 wording is the weak link. The stronger assurance comes from HIPAA with executable BAAs and GDPR compliance, backed by a trust center with reports. Sourcedocs.extend.ai/security/compliance with the docs navigationread 2026-09-12 |
|
| Observability & Auditability | Full |
|
Reviewing Workflow Run gives a record at run level. Inspecting an individual run of a pipeline shows what the agent did, which output quality reporting alone does not. Workflow Versioning sits beside it, so a run can be read against the configuration that produced it, which is what makes a run record usable months later. Evaluation adds Publishing Processors, a promotion step, so what changed and when is itself recorded. The quality layer adds field level and document level accuracy reports, confidence scores per extraction, and regression detection across evaluation sets. Extend runs a public status page, and webhook events notify a customer's own systems of run outcomes instead of polling. Extend does not say what a run record contains, such as step level timing, inputs, model version, retention or export. Sourcedocs.extend.ai workflows and evaluation navigationread 2026-09-12 |
|
| Memory & State Persistence | Partial |
|
A memory feature sits under Optimization beside Composer, the background agent that refines extraction schemas across runs. Placed with the features for improving over time, it is retained state, not a session buffer. Extend does not spell out its scope, lifetime or access rules. The system appears to write the state from its own runs, since Composer is a background agent that experiments with prompt and schema variants and refines them. User roles and permissions and per deployment isolation exist, but the scope of memory itself is not given. There is no lifetime or purge path for memory. Data handling rules cover storage, retention and deletion in general. Extend does not say whether the API, SDKs, CLI or MCP server expose the memory object. An agent written store that carries across runs goes beyond a database doing the job of memory. Sourcedocs.extend.ai Optimization navigation with extend.ai/resources Composer writingread 2026-09-12 |
|
| Deployment & Data Residency | Full |
|
Deployments run in three named regions, us1, us2 and eu1, with deployments covered in the API reference, so the region is a parameter a customer sets. The EU region is specified to the datacenter. eu1 stores and processes customer document data within the EU, with primary infrastructure in Frankfurt on AWS eu-central-1. Naming the cloud region, not just the country, is what a procurement team bound by GDPR asks for. Extend is candid about the limit. Operational telemetry that does not include document content, such as error and performance monitoring, may be processed by vendors outside the EU. Separating document data from telemetry and saying which leaves the region marks a real boundary. Extend also refers to deployment models beyond the three regions, and HIPAA compliant processing is available across all of them. Extend does not say whether those models include a self hosted or private cloud tier. Sourcedocs.extend.ai/security/complianceread 2026-09-12 |
|
| Prebuilt Agents, Templates & Packs | Not documented |
|
Generation is the opposite of a catalog, and it is Extend's deliberate position. Composer takes a few of the customer's own sample documents, generates the extraction schema and refines it in background optimization loops. The unit a customer ends up with is built for their documents, not selected from a library, and Extend sells that as the advantage on the messy documents the product exists for. There is no browsable set of units to adopt. The five APIs are product surfaces. Workflow steps are primitives a customer assembles, processors under Evaluation are the customer's own configurations being promoted, and UI Components are embeddable interface elements. None is a working unit a customer selects and runs on day one. No industry or document type packs appear, the most conspicuous absence given the product. Extend serves finance, healthcare and logistics and names Brex, Square, Checkr and Flatiron Health as customers. An invoice, claims or bank statement pack would be the obvious thing to ship, and none is offered. Supported File Types is a compatibility list. There is no schema gallery, document type template library or shareable processor, and the dashboard at dashboard.extend.ai sits behind login. Sourcedocs.extend.ai overview and full navigation, with extend.ai/resources document processing APIs writingread 2026-09-12 |
|
| Triggers & Channel Coverage | Partial |
|
Webhooks cover events, with configuration, an event catalog and best practices. Delivery is configurable, so a customer's systems are notified when a run completes or fails instead of polling. Async and batch processing are how long running document work is invoked at volume. The caller submits and is called back. A GitHub App integration under Workflows is a second invocation path, with a commit or pull request event in the customer's repository driving a document pipeline. Invocation surfaces are several but developer facing. They are the REST API, the CLI, the MCP server, the Studio dashboard at dashboard.extend.ai, the GitHub App, and embeddable UI Components that put Extend's review interface inside a customer's own product, so a reviewer working in the customer's application reaches Extend without knowing it. There is no schedule, such as cron, a recurring run or a scheduled batch. The customer's systems invoke Extend, which does not wake itself, a coherent design for infrastructure. There is also no Slack, Teams or email channel, because the customer is an engineering team. Sourcedocs.extend.ai Webhooks, General and Workflows navigationread 2026-09-12 |
|
| Model Flexibility & Routing | Partial |
|
Extend routes across more than one model. It runs an agentic OCR pipeline with intelligent routing through vision AI and custom-trained VLMs, an ensemble of frontier models combined with proprietary context engineering, and a Review Agent that uses vision language models to check low confidence extractions. The platform runs several model classes and chooses between them per document. Customer control exists over mode and version, not provider. Multiple performance modes, tuned for low latency, bulk cost efficiency or maximum accuracy, are toggled per job. Model versioning is a top level part of the product, so customers manage model change. Pinning a version is real control over which model runs. Control of the model itself stays with Extend. A customer can pin Extend's version and pick a speed, cost and accuracy profile, but cannot nominate a provider, bring their own model or route a step to a named external LLM. The models are Extend's own, including custom trained VLMs, and accuracy claims in the mid to high nineties rest on Extend controlling the ensemble. Sourcedocs.extend.ai primary navigation with extend.ai/resources document processing APIs writingread 2026-09-12 |
|
| APIs, SDKs & MCP Extensibility | Full |
|
Extend offers four developer tools, SDKs, a CLI, an MCP Server and UI Components. An MCP server, SDKs, a CLI and a REST API together make an unusually wide external surface. The API has full reference documentation, covering authentication, deployments for region selection, and webhooks with configuration, events and best practices, in a versioned documentation build stamped v2026-02-09. Rate limits, async processing and batch processing are each covered. The documentation is built for machine readers as well, with an index for AI agents at /llms.txt, a page level index at any URL with /llms.txt appended, and a Markdown version of any page with .md appended. A GitHub App integration works with Workflows, and UI Components let a customer embed Extend's review interface in its own product, an extension surface few vendors offer. The MCP server exposes Extend to a customer's assistant, while the External Data Validation step reaches the other way, calling a customer's own systems mid workflow. Extend names the SDK, CLI and MCP server but does not list their tools one by one. Sourcedocs.extend.ai Dev Tools navigation, API Reference and Webhooks sectionsread 2026-09-12 |
|
| Testing, Debugging & Optimization | Full |
|
Evaluation is part of the product, not a claim. It covers processors, publishing processors, creating and running evaluation sets, and calculating array accuracy. Evaluation sets are the core artifact. A customer assembles documents with known correct answers, runs a configuration against them and reads a score. This puts a change under test with a readable, comparable result. Publishing processors is the gate that separates this from a metrics dashboard. A processor is a configuration and publishing is a promotion step, so a change is measured before it reaches production instead of monitored after. Workflow Versioning gives the rollback side, and a test environment gives a separate place to run against. The results include field level and document level accuracy reports, regression detection across runs, confidence scores per extraction, and guidance on calculating array accuracy, how to score an extraction that returns a list, which is the hard case in document extraction. Extend adds LLM-as-a-judge scoring alongside automated accuracy reports. The optimization loop is controlled. Composer, a background agent, experiments with prompt and schema variants and refines them, measured against the evaluation sets instead of left to drift. A self improving loop with scoring the customer can read is different from absorbed learning. Sourcedocs.extend.ai Evaluation and Optimization navigation, with extend.ai/resources document processing APIs writingread 2026-09-12 |
|
| Browser & Computer Use | Not documented |
|
Every path in is an API call, an SDK, the CLI, the MCP server, a webhook or an upload, and every path out is structured data, markdown or a filled document returned to the caller. There is no interface for an agent to drive because the input is a file, so how often it breaks when interface elements change is not a question that applies. There is no hosted or local browser, desktop session, remote computer control, RPA, recorder or extension. The Dev Tools section's UI Components are interface elements Extend ships for a customer to embed, the reverse relationship. Two adjacent items fall short of computer use. The Editing capability, with Detect Form, detects fields in a document and fills them. This is form field detection inside a file, not an agent operating a form in a browser. Vision language models and an agentic OCR pipeline read the pixels of a scanned page, which is ingestion, not computer use. A model that can read a degraded scan is not a model that can click a button. Extend is infrastructure other products build on, integrated by API precisely so that no screen is involved. Sourcedocs.extend.ai full navigation including Capabilities, Dev Tools and Editing sectionsread 2026-09-12 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Product updates
Operator-1, an autonomous document extraction agent, launched in the Extend dashboard and API on all plans at a fixed 5 credits per page. The agent works in a dedicated workspace with inspection, parsing, code execution and validation tools, tries again when a check fails, and keeps the code and approach that worked for a document type as memory and runbooks within the customer's tenant. Extend reports scores of 99.3% on LongArray-Extract, 99.88% on LongExtractionBench and 96.57% on ExtractBench.
Bears on: Memory / state
View sourceExtend released an upgrade to its workflow orchestration layer that allows users to build complete document processing workflows in natural language using its Composer agent. The update also enables teams to validate these workflows with code or semantic rules and manage deployment directly through GitHub.
Bears on: Workflow orchestration
View sourcePricing
Free to start on Pay As You Go, with 10,000 credits included and $0.0125 per additional credit. Scale is $500 a month with 50,000 credits a month and $0.01 per additional credit. Enterprise is custom.
Credits, Extend's universal billing unit. Each API (Parse, Extract, Split, Classify or Edit) consumes a base number of credits per page, and extraction costs 3 credits per page plus 2 credits per page for the automatic parse, so a 10 page document costs 50 credits. The credit rules are documented at docs.extend.ai/general/how-credits-work, beside Rate Limits, Async Processing and Batch Processing.
Variable cost rationale
Extend meters credits per page processed, and each API consumes its own base number of credits per page. Extraction costs 3 credits per page plus 2 for the automatic parse. Cost therefore tracks how many pages a customer processes and which of the five APIs a workflow calls. For the named customers, Brex, Square, Checkr and Flatiron Health processing millions of documents, that volume follows their own business rather than a headcount decision. Nothing published says whether Composer's background optimization loops, which experiment with prompt and schema variants, draw credits. Pay As You Go has no contract or monthly minimum. Scale's $500 a month includes 50,000 credits a month and lowers the rate for additional credits to $0.01, against $0.0125 on Pay As You Go, and Enterprise sets custom per credit pricing. Rate limits are documented on their own page and act as a ceiling, with higher limits on Scale and custom limits on Enterprise. The evaluation framework lets a team measure cost against accuracy before scaling.
Sales call required
Mixed (some tiers require a call)
Free / trial
Pay As You Go is free to start, with 10,000 credits included
Lowest paid plan
Pay As You Go at $0.0125 per credit after the first 10,000, or Scale at $500 per month with 50,000 credits
Key ambiguities
The price of a document depends on the APIs a workflow calls, because each API consumes its own base number of credits per page. Extraction costs 3 credits per page plus 2 credits per page for the automatic parse, and nothing published says whether Composer's background optimization loops also draw credits. docs.extend.ai has a How Credits Work page under General, beside Rate Limits, Async Processing and Batch Processing, so credits are the billing unit and rate limits are a documented constraint. That is a usage model, not a seat model. Both motions exist. The documentation links a Studio dashboard at dashboard.extend.ai and a log in, alongside a Book a demo path to a named salesperson's scheduling link, so self serve entry and a sales led enterprise motion run in parallel. Enterprise adds BYOC (your VPC) and hybrid deployments, SSO and SAML, advanced RBAC and custom rate limits.
Missing data
Enterprise per credit pricing is custom and unpublished, and the size of the volume discounts on Scale is not stated. extend.ai/pricing gives no rate limit figures for any plan and does not say whether unused Scale credits roll over. No minimum or contract term is published for Scale or Enterprise.
Related vendors
- AgentOps — Agent observability and debugging platform: open source SDKs trace…
- AgentPhone — AgentPhone gives the AI agents a team builds their own US or…
- Agno — Python agent framework and AgentOS runtime (formerly Phidata) for…
- AIsa — Resource and payment gateway for AI agents, with one key to 110+…
- Albato — Albato is a no code integration platform that sells an embedded,…
- AlphaBitCore — AI control plane for regulated financial firms: one gateway enforces…
Alternatives to Extend
The closest documented capability profiles to Extend among agent infrastructure platforms tracked by Agentic Index, ordered by similarity on the same 14 point evidence the rankings use. No vendor pays for placement.
- Inngest10.0 / 14Adds documented Prebuilt Agents, Templates & Packs
- Bernstein10.5 / 14Adds documented Prebuilt Agents, Templates & Packs
- C TWO8.5 / 14Fuller documented coverage on Integrations & Tool Calling and Triggers & Channel Coverage
- DBOS8.5 / 14Adds documented Prebuilt Agents, Templates & Packs
- Fiddler AI8.5 / 14Fuller documented coverage on Triggers & Channel Coverage and Model Flexibility & Routing
- Hatchet8.5 / 14Adds documented Prebuilt Agents, Templates & Packs
Similarity is computed from each vendor's Agentic Index coverage score evidence, axis by axis, not from the totals. How this evidence is graded