Open OCR CLI

Agent-first, provider-neutral document extraction for the command line.

Turn images, PDFs, and public URLs into text or validated structured data — through Gemini, Kimi K3, Meta Muse Spark, OpenRouter, or any OpenAI-compatible endpoint — behind one consistent extraction contract.

Quick start · Modes · Commands · Agents and MCP · Providers · Full CLI reference

Quick start

Requires Node.js 20.19+, 22.13+, or 24+ and one provider key (--dry-run needs none).

npm install --global open-ocr-cli

export GEMINI_API_KEY="your-key"
open-ocr-cli extract invoice.pdf

Switch providers with a flag:

open-ocr-cli extract invoice.pdf --provider kimi          # MOONSHOT_API_KEY
open-ocr-cli extract invoice.pdf --provider openrouter \
  --model moonshotai/kimi-k3                              # OPENROUTER_API_KEY

Keys are read only from the environment or a project .env — never from a flag or config file. With no arguments the CLI prints help and never prompts, so scripts and CI stay deterministic; interactive, init, and doctor cover the guided paths.

Also available as a Docker image (ghcr.io/cyanxxy/open-ocr-cli), a GitHub Action (cyanxxy/open-ocr-cli@v3), and a Homebrew formula.

Extraction modes

Schemas are validated locally — both the schema and the model's returned value — before anything is written. Invalid JSON or a schema mismatch is never persisted as a success.

Agentic mode uses a two-loop design: an outer document loop evaluates confidence and coverage, while an inner loop lets the model call one tool (structure analysis, batch field extraction, region re-OCR), inspect the result, and decide what to do next. The document is sent once.

Commands

# Recursive, resumable batch
open-ocr-cli extract ./documents --output ./results --concurrency 4 --resume

# Structured invoice artifacts (markdown + JSON + CSV)
open-ocr-cli extract ./invoices --preset invoice --format all --output ./invoice-results

# Validate the full plan without credentials, API calls, or writes
open-ocr-cli extract ./documents --dry-run

Extracted content and JSONL events go to stdout; progress and diagnostics go to stderr — every command composes safely in a pipeline.

Flags, configuration keys, batch semantics, cost controls, and exit codes: packages/cli/README.md

Agents and MCP

Three entry points wrap the same job engine: extract (human-facing), run (versioned protocol), and mcp (Model Context Protocol).

Machine protocol — run executes a protocol v2 request validated against published Draft 2020-12 JSON Schemas, returning a typed result object or an ordered JSONL event stream with stable sequence numbers and typed error codes carrying recovery hints:

open-ocr-cli capabilities --json
open-ocr-cli run --request request.json --response-format jsonl

MCP server — open-ocr-cli mcp starts a stdio server exposing ocr_extract, ocr_run_agentic, and ocr_web, plus an open-ocr://capabilities resource. The host must open MCP revision 2026-07-28; clients that use the earlier initialize handshake are rejected:

{
  "mcpServers": {
    "open-ocr": {
      "command": "open-ocr-cli",
      "args": ["mcp"]
    }
  }
}

MCP local inputs use the same typed { "type": "path", "path": "…" } objects as the machine protocol. Partial document results carry a typed partialReason and nextAction; the separate extract --jsonl v1 dialect is published through open-ocr-cli schema jsonl-v1.

Agent skill — a validated skill for Claude Code, Codex, and compatible agents ships at integrations/open-ocr/skills/open-ocr/SKILL.md and in the npm package under skills/open-ocr/.

All three are reference-first: large bodies land in .open-ocr-results/<runId> artifacts instead of flooding an agent's context, cancellation is honored, and under mcp stdout carries only transport messages.

Providers

Named profiles supply endpoints, credential variable names, and multimodal wire formats; openrouter and openai-compatible take arbitrary upstream model IDs. Any provider can also run through Cloudflare AI Gateway with --gateway cloudflare, including BYOK.

Model catalogs, reasoning levels, cost estimates, and gateway configuration are documented in the CLI reference.

Limits and guarantees

Every discovered input gets a succeeded, partial, failed, or skipped result — no silent loss. Batches are fingerprinted and resumable, output is lock-guarded and no-clobber, and truncated, blocked, or schema-invalid model responses fail closed rather than being written as success.

There is no backend and no telemetry: documents go to the selected provider or gateway and nowhere else. Web OCR rejects credentials in URLs, localhost, private ranges, and tunnel hosts, and pins DNS across redirects. Report vulnerabilities privately via SECURITY.md.

Repository layout

One npm workspace, two packages over a shared extraction engine:

Both packages are typechecked without DOM libraries on purpose: a browser API reaching this code fails the build rather than failing at runtime. Anything host-specific — such as region cropping — enters the engine through an adapter the host supplies (packages/cli/src/nodeRegionCropper.ts).

Contributing

Setup, PR gates, and release steps are in CONTRIBUTING.md. Behavior changes need tests or evaluation evidence; the evaluation suite is documented in evals/README.md.

Issues · Releases · Changelog · Security · Code of Conduct

MIT Licensed — see LICENSE