Skip to main content

ezdgn

ezdgn is a native, read-only V8 and read/write V7 DGN toolkit for Python, implemented with an independently written Rust core and PyO3 bindings. It has no ODA SDK, binary, or runtime dependency.

Installation

ezdgn requires Python 3.10 or newer.

python -m pip install ezdgn

Install the optional Matplotlib renderer when preview images are needed:

python -m pip install "ezdgn[plot]"

Building from source requires Rust 1.83 or newer. The extension uses Python's stable ABI (abi3) with a Python 3.10 minimum.

Supported scope

Format and operation Support
V7/ISFF 2D read Native entities, hierarchy, metadata, linkages, and raw records
V7/ISFF 2D write Standalone or custom-seed creation of common primitives
V7/ISFF 3D Signature, raw record, and common-header inspection only
V8 DGN read Native 2D/3D models, common geometry, hierarchy, linkages, auxiliary data, and lossless raw objects
V8 DGN write/edit Not supported

The V7 reader decodes line, line string, shape, curve, ellipse, arc, text, cell, text node, complex chain/shape, and B-spline records as native entities. It restores parent/child relationships without flattening component records, decodes typed attribute linkages while retaining their exact bytes, applies high-precision sub-UOR coordinate corrections alongside the stored integers, and resolves outline/fill colors through the active color table. Every record, including unsupported control and application elements, retains its original bytes.

The writer creates standalone V7 2D files from a bundled empty seed. A custom seed can be supplied to preserve its units, origin, and design plane. It writes line, line string, shape, curve, ellipse, arc, circle-as-ellipse, and raw-byte text entities with basic symbology and shape fill linkage. Coordinates outside the selected seed's design plane are rejected instead of silently clipped.

The native V8 reader decodes model metadata, line, line string, shape, curve, point string, text, ellipse, arc, text node, complex chain/shape, cell, shared-cell instance, B-spline curve/poles, and a bounded dimension anchor. Unknown graphical, control, name-table, linkage, and auxiliary records remain addressable with exact bytes. Product-specific custom-object semantics, full dimension annotation semantics, V8 writing/editing, and 3D screen projection for plotting remain outside the supported scope.

Real-world files are read tolerantly where the damage is local. Deleted V7 records (the header's deleted bit) never reject a file: MicroStation extends a deleted complex header's words-to-follow over its components, so deleted headers are not counted against their declared component number. A live complex header that encloses no components is demoted to an unsupported element. In V8, a single element with malformed geometry (for example a line string with one vertex) becomes an unknown element instead of rejecting its model, and an attribute boundary past the end of its object is read as "no attributes". The per-object size limit defaults to 16 MiB; objects are slices of pages that are already bounded by the 64 MiB per-page and 1 GiB total inflation limits, which remain the decompression guards.

V7 3D geometry is not supported. The V7 writer does not yet create cells, complex elements, B-splines, arbitrary linkages, or perform in-place editing. The raw V7 record framing is shared by 2D and 3D files, so scan_records() can inspect a 3D stream safely without implying V7 3D entity support.

2D entity API

import ezdgn

drawing = ezdgn.readfile("drawing.dgn")

# All records remain ordered and lossless. entities contains only top-level
# graphics; all_entities also includes drawable component records.
print(len(drawing.elements), len(drawing.entities), len(drawing.all_entities))

for entity in drawing:
    print(entity.dxftype(), entity.level, entity.style)

for line in drawing.query("LINE"):
    print(line.start_uor, line.end_uor)
    print(line.start_master, line.end_master)

for text in drawing.query("TEXT"):
    print(text.text_bytes)
    print(text.decode_text())  # ASCII with strict errors by default
    print(text.decode_text("cp932", errors="replace"))

for cell in drawing.query("CELL"):
    print(cell.name, cell.origin_master, cell.transform)
    for component in drawing.children(cell):
        print("  ", component.dxftype())

for element in drawing.elements:
    for linkage in element.linkages:
        print(linkage.kind, linkage.linkage_type_name, linkage.raw_bytes)

drawing.entities contains supported top-level drawable entities only. drawing.all_entities is a flat tuple in original drawing.elements order: it includes both Cell/ComplexChain/ComplexShape container headers and their supported drawable descendants at every depth. It excludes control records, UnsupportedElement instances, and non-drawable B-spline support records; those remain available through the lossless drawing.elements sequence. drawing.unsupported_elements provides a filtered diagnostic view for unknown kinds. Use parent(), children(), or descendants() when a tree view is needed.

Ellipse, arc, curve, and B-spline entities retain their native parameters or control records; they are not flattened to polylines. parent_index and child_indices refer to the lossless drawing.elements sequence, while drawing.parent(), children(), and descendants() resolve the objects. Stored integer UOR, sub-UOR-corrected floating UOR, and optional master-unit coordinates coexist. drawing.color_table is the last type-5, level-1 color table in file order. drawing.resolve_color(), entity.style.rgb, and fill_rgb use that table when present and otherwise fall back to the standard MicroStation V7 256-color palette.

V7 DGN does not record a text code page. Text.decode_text() therefore uses encoding="ascii", errors="strict" as a deterministic default; pass the project encoding explicitly for non-ASCII text.

Known DMRS/database, association ID, shape fill, and high-precision linkages have typed fields. Unknown user linkages and malformed trailing attribute bytes remain accessible through read-only raw views. Shared-cell definition/instance types 34/35 remain raw because the public ISFF chapter does not specify their layout. A display header whose attribute pointer does not land inside the record (seen in vendor element type 56 blocks from real designs) no longer fails the whole file: the element keeps its geometry and simply carries no linkages. Text nodes whose writer stored the text strings inside the node record (the node's words-to-follow spans the whole complex group) are split by the record scanner into the node header followed by ordinary text records. B-spline headers whose components do not match the header are kept as raw records instead of rejecting the design.

3D V7 designs are read into the same 2D model by projecting onto the XY plane: coordinates drop Z, and elements that carry an orientation quaternion (text, text nodes, ellipses, arcs, cells) are rewritten as the equivalent 2D element seen from above — a tilted circle becomes an ellipse with the correct principal axes, and arc start/sweep angles are remapped into that ellipse's parameter space. DesignSettings.dimension still reports 3, so consumers can flag the projection. Shared cells and high-precision linkages are not decoded for 3D designs yet.

The high-level read()/readfile() API deliberately rejects V7 3D files. scan_records() and inspect_headers() still support bounded inspection of their shared record framing and metadata.

Native V8 API

Use open_document() when a caller may receive either format. The existing read() and readfile() names remain V7-specific for compatibility.

import ezdgn

document = ezdgn.open_document("drawing.dgn")
if isinstance(document, ezdgn.V8Document):
    for model in document.models:
        print(
            model.metadata.name,
            model.metadata.dimension,
            model.metadata.master_unit,
        )
        for entity in model.entities:
            print(entity.dxftype(), entity.level, entity.common.color_index)

        for text in model.query("TEXT"):
            print(text.data.text, text.data.text_bytes)

        for cell in model.query("CELL"):
            for component in model.children(cell):
                print("  ", component.dxftype())

V8Model.elements is the complete graphical-object sequence. V8Model.entities is the feature-oriented view: complex/cell headers remain native aggregate entities, while TextNode headers yield their text children. parent(), children(), and descendants() navigate the original hierarchy. 2D and 3D coordinates coexist as (x, y, z) tuples; no Z coordinate is dropped. For complex headers, common.stored_dimension retains the header bit and common.dimension reports the effective dimension inherited from children. V8 text exposes its stored width_multiplier_raw and height_multiplier_raw alongside corrected width_uor, height_uor, width_master, and height_master distances.

The raw scanner is available independently of semantic decoding:

raw = ezdgn.scan_v8_objects("drawing.dgn")
for model in raw.models:
    for obj in model.graphical_objects:
        print(
            obj.stream_path,
            obj.inflated_offset,
            obj.element_type,
            obj.role,
            obj.raw_bytes,
        )

Every extraction and decode stage is bounded. Override the defaults as one immutable policy object when processing untrusted files:

limits = ezdgn.V8ScanLimits(
    max_file_size=256 * 1024 * 1024,
    max_objects=250_000,
    max_total_inflated_bytes=512 * 1024 * 1024,
)
document = ezdgn.read_v8("drawing-v8.dgn", limits=limits)

The level table of a V8 file is available as document.levels (one V8Level per defined level: level_id as graphical headers store it, the user-visible number, name, and the level's own line style, weight and color index) and document.level(level_id). An element header that holds the ByLevel sentinel 0xffffffff in its weight, color or style takes the value of its level.

Malformed structure raises InvalidDgnError; configured resource ceilings raise DgnLimitError. The complete clean-room boundary, evidence ledger, and known limitations are recorded in docs/v8/SCOPE.md, docs/v8/FORMAT_NOTES.md, and docs/v8/PROVENANCE.md.

Plotting parsed drawings

The optional renderer can display a parsed V7 2D drawing or a V8 document and save it as an image without changing the native entity model:

import ezdgn

drawing = ezdgn.readfile("drawing.dgn")

figure, axes = ezdgn.plot(
    drawing,
    text_encoding="cp932",
    background="#111111",
)
figure.savefig("preview.png", dpi=150, bbox_inches="tight")

# Or render and save in one call.
ezdgn.save_plot(drawing, "preview.png", text_encoding="cp932")

# V8 uses already decoded native text and preserves Z in the object model;
# the preview is an explicit XY projection.
ezdgn.save_plot(ezdgn.read_v8("drawing-v8.dgn"), "preview-v8.png")

The equivalent CLI command is:

ezdgn plot drawing.dgn -o preview.png --encoding cp932

Omit -o to open an interactive Matplotlib window. Use --monochrome for a high-contrast preview, --hide-text to suppress text, or --coordinate-space uor when master-unit coordinates are unavailable. Run ezdgn plot --help for the complete option list.

Lines, line strings, shapes, ellipses, arcs, text, and drawable components of cells and complex elements are rendered. Ellipses and arcs are sampled only for display. Native type-11 curves and B-spline curves are previewed from their parsed control sequences; the source records and entity parameters are never flattened or modified. V8 3D geometry uses an XY preview projection, and orientation matrices are retained but not applied by the 2D renderer. V7 text does not store its code page, so the caller must select the correct encoding for non-ASCII text. Geometry and text with compatible display styles are batched to keep large previews practical.

V7 writer and optional custom seeds

import ezdgn

doc = ezdgn.new()  # uses the bundled empty V7 2D seed
msp = doc.modelspace()

msp.add_line(
    (0, 0),
    (10, 5),
    dgnattribs={"level": 2, "color": 3, "line_weight": 2},
)
msp.add_line_string([(0, 10), (5, 15), (10, 10)])
msp.add_shape(
    [(20, 0), (30, 0), (30, 10), (20, 10)],
    fill_color=6,
)
msp.add_ellipse((25, 25), primary_axis=5, secondary_axis=3, rotation=30)
msp.add_arc(
    (40, 5),
    primary_axis=5,
    secondary_axis=3,
    start_angle=30,
    sweep_angle=120,
)
msp.add_text("日本語", (0, 30), height=2, encoding="cp932")

doc.saveas("drawing.dgn")
roundtrip = doc.readback()

Pass a custom seed as ezdgn.new("project_seed.dgn") to use its units, origin, design plane, and active color table. new() copies the mandatory TCB, digitizer setup, level symbology, and last active color table from the selected seed. Set copy_seed_elements=True to retain every seed record, including any graphics. Text encoding is not recorded by V7 DGN, so add_text() accepts bytes directly or uses its encoding="ascii", errors="strict" defaults for str input.

Raw record API

import ezdgn

info = ezdgn.detect_format("drawing.dgn")
print(info.kind, info.dimension)

scan = ezdgn.scan_records("drawing.dgn")
print(len(scan.records), scan.termination)

for record in scan.records:
    print(record.offset, record.element_type, record.level, record.raw_bytes)

The V8_CFB result means that the input has the generic CFB signature used by V8 DGN files. It is intentionally described as a candidate because the outer signature alone does not prove that DGN-specific streams are present.

The bounded container inspector is the structural-only entry point. It verifies the known DGN root markers without decoding DGN stream contents:

container = ezdgn.inspect_v8_container("drawing-v8.dgn")
print(container.has_dgn_v8_markers)
print(container.model_storage_paths)
for entry in container.entries:
    print(entry.path, entry.kind, entry.size_bytes)

inspect_v8_container() alone is structural identification, not an entity or fidelity claim. Use scan_v8_objects() for bounded, exact raw objects and read_v8() or open_document() for native semantic decoding. The legacy read(), readfile(), and scan_records() contracts remain V7-specific and therefore reject V8 input. No V8 path silently converts to V7 or flattens native objects.

Design settings and common headers

import ezdgn

headers = ezdgn.inspect_headers("drawing.dgn")
settings = headers.design_settings

print(settings.master_unit_name, settings.uor_per_master)
print(settings.global_origin_master)

for element in headers.elements:
    common = element.common_header
    if common is not None:
        print(
            element.record.element_type,
            common.range.low_master,
            common.range.high_master,
            common.symbology.color,
        )

read_design_settings() decodes only the leading TCB. inspect_headers() pairs every raw record with its standard common header when that element type has one. Attribute bytes remain available as a read-only zero-copy ElementMetadata.attribute_view.

The same inspection is available from the CLI:

ezdgn inspect drawing.dgn
ezdgn inspect drawing.dgn --records --json
ezdgn inspect drawing.dgn --headers --json
ezdgn inspect drawing.dgn --entities --json
ezdgn inspect drawing-v8.dgn --json

Development

python -m venv .venv
. .venv/bin/activate
python -m pip install "maturin>=1.13,<2" "pytest>=8" "matplotlib>=3.8"
maturin develop
cargo fmt --check
cargo clippy --workspace --all-targets --all-features -- -D warnings
cargo test --workspace --all-features
python -m pytest
sha256sum -c tests/data/dgn/SHA256SUMS
uv lock --check

Run the bounded V8 fuzz target and parser microbenchmark separately:

cargo +nightly fuzz run v8_read -- -max_len=16777216
EZDGN_BENCH_ITERATIONS=100 cargo bench -p ezdgn-core --bench v8_read

Build a distributable wheel with:

maturin build --release --out dist

License

ezdgn is released under the MIT License. The bundled empty seed and test fixtures retain the separate upstream terms documented in src/ezdgn/_data/README.md and tests/data/dgn/README.md.

Metadata

Release files for ezdgn 0.2.9

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for ezdgn 0.2.9
File Size Uploaded
ezdgn-0.2.9.tar.gz 146.0 kB Details

Built distributions (wheels)

Table of built distributions (wheels) for ezdgn 0.2.9
File
ezdgn-0.2.9-cp310-abi3-win_amd64.whl CPython 3.10 abi3 Windows x86-64 Details
ezdgn-0.2.9-cp310-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl CPython 3.10 abi3 Linux glibc 2.17+ x86-64 Details
ezdgn-0.2.9-cp310-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl CPython 3.10 abi3 Linux glibc 2.17+ ARM64 Details
ezdgn-0.2.9-cp310-abi3-macosx_10_12_x86_64.macosx_11_0_arm64.macosx_10_12_universal2.whl CPython 3.10 abi3 macOS 10.12+ x86-64, macOS 11.0+ ARM64, macOS 10.12+ universal2 (ARM64, x86-64) Details

Total release size: 2.7 MB

Release files / ezdgn-0.2.9.tar.gz

Download URL ezdgn-0.2.9.tar.gz
Size 146.0 kB
Tags Source
SHA-256 checksum
How to use checksums
dfa40246c2eb6918d32d005f05b3777ae43dba59780863dd26a6dbfe1a623b31
BLAKE2b-256 checksum
How to use checksums
d402b0e6056110200c8511a16789c14cf906c89e207194fc22ec53d53b7785a1
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Oct 15, 2026.

Transparency log

Release files / ezdgn-0.2.9-cp310-abi3-win_amd64.whl

Download URL ezdgn-0.2.9-cp310-abi3-win_amd64.whl
Size 435.1 kB
Tags CPython 3.10 Windows x86-64 abi3
SHA-256 checksum
How to use checksums
841f4d051ac30cbfd5781391cad29a6fdf4de14fe59db736e2384e1228d67fb0
BLAKE2b-256 checksum
How to use checksums
2a8e4b6048c50c16a0be5b7ecccf004365a30fe87fcec7a6bdd181fcea377565
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Oct 15, 2026.

Transparency log

Release files / ezdgn-0.2.9-cp310-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl

Download URL ezdgn-0.2.9-cp310-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
Size 586.9 kB
Tags CPython 3.10 Linux glibc 2.17+ x86-64 abi3
SHA-256 checksum
How to use checksums
9198d15dbe73da05c9383dc44f25d3a22ae7990239494d0ce35ca3d90c7a7fce
BLAKE2b-256 checksum
How to use checksums
df75335441a315251a513a21f10f2a7c7da4e761d91e6abe412f473cf6e55a7f
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Oct 15, 2026.

Transparency log

Release files / ezdgn-0.2.9-cp310-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl

Download URL ezdgn-0.2.9-cp310-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl
Size 575.4 kB
Tags CPython 3.10 Linux glibc 2.17+ ARM64 abi3
SHA-256 checksum
How to use checksums
1cead90a47669d95d4aa1e7575d6bb034140c840db34651ce3b5a4fde2ac9ccb
BLAKE2b-256 checksum
How to use checksums
192fbb4e405accefeb7d5d9f245d73c38de46202069b95d3c76ee7692c2e549b
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Oct 15, 2026.

Transparency log

Release files / ezdgn-0.2.9-cp310-abi3-macosx_10_12_x86_64.macosx_11_0_arm64.macosx_10_12_universal2.whl

Download URL ezdgn-0.2.9-cp310-abi3-macosx_10_12_x86_64.macosx_11_0_arm64.macosx_10_12_universal2.whl
Size 1.0 MB
Tags CPython 3.10 abi3 macOS 10.12+ universal2 (ARM64, x86-64) macOS 10.12+ x86-64 macOS 11.0+ ARM64
SHA-256 checksum
How to use checksums
a24453396b8103693c6f5a4c8aacdb80a21eb4362abe3f599e8345805143848e
BLAKE2b-256 checksum
How to use checksums
5de2b9bf5c07b41329bada5678efaefee63046135738b0b2a851aabac14d22ad
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Oct 15, 2026.

Transparency log

Release history Release notifications | RSS feed

This release

0.2.9 This release

5 release files

0.2.8

5 release files

0.2.7

5 release files

0.2.6

5 release files

0.2.5

5 release files

0.2.4

5 release files

0.2.3

5 release files

0.2.2

5 release files

0.2.1

5 release files

0.2.0

5 release files

0.1.2

5 release files

0.1.1

5 release files

0.1.0

5 release files

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page