Workflow engine
Build document pipelines that handle ingestion, parsing, extraction, validation, branching, human review, and downstream API calls.
layer 01
MCP for document operations
Agent-Native
Document Automation
Retab makes document workflows inspectable by agents. Capture each run as a trace, expose the right MCP tools, and let agents verify, debug, and improve workflows from the source documents.
codex mcp add retab --url https://mcp.retab.com/mcpClaude Codev2.1.3
Opus 4.7
> Create a workflow to extract data from bank statements
↳workflows_create{ "name": "bank_statement_reconciliation" }
⎿✓draft created
↳workflows_blocks_create{ "type": "extract+function" }
⎿✓extract_transactions + balance_check configured
↳workflows_runs_create{ "version": "draft" }
⎿✓completed · 143 transactions · balance_check passed
●Workflow drafted: PDF -> extract fields + transactions -> balance check -> clean JSON.
Run system of record
Every run preserves the original files, intermediate outputs, field values, validation results, transformations, and final payloads.
layer 02
MCP toolbox
Agents can query runs, inspect source regions, search document text, retrieve provenance, diagnose workflows, and review experiments.
layer 03
From outputs to operations
Document workflows, grounded in the source of truth
Extraction quality has improved, but most pipelines still stop at a final JSON response. Retab records how that response was produced, so failures can be investigated inside the same system that generated them.
“The workflow becomes part of the agent's working memory. Just as coding agents only became useful once they could see the codebase, operational agents need direct access to the systems where the truth lives.”
Black-box extraction
PDF
JSON
When a value is wrong, the system cannot explain which step produced it, where it came from, or what should happen next.
Retab workflow record
Sources bounding boxes
parsed text
extracted fields
validation checks
step executions
provenance
A run becomes an inspectable object that humans and agents can query, debug, correct, and improve over time.
Agent-ready context
Give agents the tools operators already use
The last mile of automation is not another extractor. It is the ability to verify, debug, and decide what happens when documents are ambiguous, incomplete, or internally inconsistent.
files_list
Discover available files and workflow inputs.
files_grep
Find exact values or phrases in source content.
files_inspect
Open the native context around an anchor.
files_get_sources
Trace extracted fields back to evidence.
workflows_steps_get
Read full step inputs and outputs.
workflows_diagnose
Find structural, config, and review issues.
Agent in the loop
Humans become the fallback, not the first line of defense
Route uncertain outputs to agents that can inspect the run, trace inconsistencies, cross-check the source document, and decide whether to correct, retry, or escalate.
step 01
Failed check
step 02
Inspect trace
step 03
Fix or retry
step 04
Escalate only if needed
Self-Improving Workflows
Automation improves without leaving the workflow
Retab keeps execution, memory, inspection, and iteration in one place. The same workflow that produces structured data can be tested, diagnosed, re-run, and optimized by agents over time.
Build
01Create workflows from document samples with extraction, validation, branching, and review.
Run
02Capture every step as a structured record tied to original files and output handles.
Inspect
03Use MCP tools to search, retrieve sources, inspect pages, and diagnose failed checks.
Improve
04Execute block changes, run evals and experiments, then publish better workflows.
Build AI native document processing infrastructure
Connect Retab MCP to Claude, Codex, or any compatible client and give agents direct access to document automation infrastructure.