Releases · cyanxxy/open-ocr-cli

Release list

v2.7.0 Latest

Latest

github-actions released this 29 Jul 14:26

v2.7.0

82423af

2.7.0 - 2026-07-29

Added

MCP now serves both the 2025-11-25 and 2026-07-28 protocol revisions from
one stdio entry, including modern discovery and cache hints.
OCR tools advertise the v2 result schema and expose written artifacts as
native resource_link content blocks.
OPEN_OCR_MCP_CONFIRM=1 requires an elicited confirmation before provider
work while leaving local dry runs ungated.

Fixed

MCP confirmations are signed, bound to the exact request, expire, and can be
used only once, preventing forged, changed, or replayed approvals.
Clients that support URL elicitation but not forms now receive the typed
configuration error instead of a protocol-level capability failure.
MCP shutdown preserves the documented 130 and 143 exit statuses for
SIGINT and SIGTERM.

Changed

Updated the split MCP client and server packages from the v2 beta to the
stable 2.0.0 release.

Full comparison

Assets 6

v2.6.0

github-actions released this 26 Jul 22:07

v2.6.0

4083c65

2.6.0 - 2026-07-26

Breaking

extract --jsonl now emits one record family. Fatal failures end the stream
with {"type":"error","version":1,...} instead of a protocol-shaped
run.failed, so every line carries "version": 1 and no protocolVersion.
Previously a strict validator rejected every line when a run succeeded and
accepted every line when it failed. run --response-format jsonl is
unchanged and remains the protocol dialect.

Fixed

Template mode produced 400 INVALID_ARGUMENT for every preset in every
earlier release. Structured output compiles an array bound into that many
copies of the item grammar, so the row cap in the wire schema exceeded a
limit the API enforces. The cap is applied after parsing instead, where it
already was, and a static check now fails the build rather than the request.
--resume aborted an entire batch when a single input changed, leaving even
unchanged documents unprocessed. A run now replaces only the artifacts its
own manifest recorded for that input, and an unrelated file at an artifact
path still conflicts.
Single-document runs into an output directory kept no manifest, so re-running
one failed against its own output. An output path naming a single file is
still left alone, and can no longer be turned into a directory by the lock.
Document failures reported an unredacted message alongside a redacted one, so
credentials could survive into JSONL records, the resume manifest, and
batch-summary.json. Both now derive from one redacted source.
Provider rejections reported the raw response body. The readable sentence is
extracted where the error is created, and the untouched payload is retained
as the cause.
A rejection naming no field is no longer blamed on the request schema unless
the schema exceeds a measured limit; unfamiliar keywords are advisory. Schema
identifiers published by capabilities now resolve in schema.
MCP reported a successful tool call when a run was partial, cancelled, or
cost-limited. isError now derives from the run envelope.
CSV asked of a preset that extracts one record per document is refused before
the credential gate rather than after a billed request. Inline delivery
combined with an output directory names both fields and the conflict.
Typed errors carry a next action on every code, and errors identified by the
code path are no longer re-classified from their message text.

Changed

Directory scans skip node_modules, dist, build, vendor, and target
by default, and report what they passed over. Use --no-default-excludes to
include them.
Directory scans no longer ignore uppercase file extensions, so a folder of
.PNG or .PDF scans is discovered and extracted.
A single-document run into a directory now takes the batch lock it needs to
keep a manifest, so simultaneous runs into one directory conflict rather than
racing. Sequential runs and single-invocation batches are unaffected.

Full comparison

Assets 6

v2.4.1

github-actions released this 26 Jul 14:35

v2.4.1

c2cb818

2.4.1 - 2026-07-26

Repository metadata

Updated the npm package repository, homepage, issue, container, and release
references to the canonical cyanxxy/open-ocr-cli repository.
Kept the private cyanxxy/gemini-ocr web application repository separate
from the public CLI package. There are no runtime behavior changes.

Full comparison

Assets 6

v2.4.0

github-actions released this 25 Jul 20:31

v2.4.0

2ed4376

2.4.0 - 2026-07-25

CLI packaging and development boundaries

Moved the CLI source, build configuration, and Node-only TypeScript project
into packages/cli so the publishable package owns its complete build path
while continuing to share the provider-neutral OCR engine.
Added dependency-sync coverage that verifies every external CLI runtime
import is declared in the published package, and tightened Node-only
typechecking so browser APIs cannot enter the command-line dependency graph.
Made the packaged Open OCR skill a generated copy of its canonical
integration source, pinned the npm toolchain used by releases, and made
TypeScript checks incremental for faster local and CI validation.

Release and dependency security

Split releases into verification and publication stages with main-branch and
tag-order checks, serialized publishing, exact-tarball smoke tests, artifact
checksums, and idempotent npm publication checks.
Added npm provenance, container SBOM and provenance attestations, immutable
container publication before latest promotion, guarded Marketplace major
tag updates, least-privilege workflow permissions, and commit-pinned GitHub
Actions.
Upgraded React Router to 8.3.0 to resolve its RSC-mode CSRF advisory, added
Dependabot coverage, and narrowed the audit exception to the remaining
transitive development-tool advisory.

Full comparison

Assets 6

v2.3.0

github-actions released this 22 Jul 13:52

v2.3.0

5f00fa4

2.3.0 - 2026-07-22

Model Context Protocol support

Added open-ocr-cli mcp, a stdio MCP server with dedicated tools for simple
and template extraction, agentic OCR, and public-URL extraction.
Added an MCP capabilities resource, structured tool results, cancellation,
progress notifications, dry runs, and reference-first artifact delivery on
top of the same validated job service used by the CLI.
Bundled the MCP server runtime and updated the packaged Open OCR agent skill
and setup documentation for Codex and other MCP clients.

Machine and Web OCR reliability

Added URL inputs and Web analysis options to protocol v2, including schema
validation, mixed-input rejection, a 20-URL request limit, and explicit
capability advertisement.
Routed machine and MCP URL work through the shared Web OCR service so dry
runs, retries, usage accounting, delivery, progress, and typed failures stay
consistent across terminal and agent integrations.
Centralized provider runtime and retry policy handling, extended retries to
Web OCR, added doctor --check-credentials, and made no-argument CLI calls
print deterministic help instead of opening an interactive prompt.
Refreshed the root and npm documentation around installation, providers,
automation, MCP setup, credential hygiene, and the explicit interactive
command.