How this compares

Most comparisons in this space go wrong by putting products from three different categories into one table. The honest first question is what you are building.

Open Integration Engine — the community fork of Mirth Connect, made after Mirth moved to a commercial-only license — and its commercial siblings are a different category of thing entirely. An engine gives you channels, routing, a management UI, transformers, persistence, retry, alerting, and an operations story. It is a system you deploy and run.

A library gives you a function call. If your problem is “forty interfaces, three hospitals, and someone has to be paged when one stops”, an engine is the right answer and no amount of crate substitutes for it.

These crates are useful alongside an engine rather than instead of it:

• The service at the end of a channel, where you would otherwise be writing the v2 parsing again in whatever language that service is in.
• A shell-level check, redaction, or normalisation, using the command-line tools, without standing anything up.
• Getting a message into a bug report. er7-redact --report prints the paths and actions with no values in the output, which is the shape a ticket can carry.

HAPI HL7v2 is the reference open-source HL7 v2 library, in Java, dual licensed under MPL 1.1 and GPL 2.0. It has been maintained for two decades, ships a generated typed model for every segment and message of every release, and has seen far more real traffic than anything here. NHapi is its .NET port; hl7apy is the best-established Python option.

If your platform is the JVM, use HAPI. That is not modesty: a twenty-year-old library with complete release coverage and a large user base is the lower-risk choice, and reimplementing it in a language you were not otherwise using is a bad trade.

The license row decides some evaluations outright. A permissive option matters if you are linking into a closed-source product; a copyleft option matters if your organisation prefers one. Offering five is how this project avoids having that conversation with anyone.

A small field, and cooperation beats competition in it. The figures below were read from the crates.io API on 2026-08-26, as context rather than as a ranking.

When to choose one of those instead

These are real reasons, not hedges.

• You want timestamp types. hl7-parser parses HL7 timestamps into chrono, time, or jiff. er7 returns the text and leaves the interpretation to you, because doing otherwise means a dependency.
• You want streaming with bounded memory over very large inputs. hl7v2_stream is built for that; er7 parses a message into a tree.
• You want Serde on the parser itself. hl7-parser has a feature flag. Here it is a separate crate, precisely so that users who do not want Serde do not pay for it.
• You are already using one and it works. Switching a working HL7 parser is rarely the highest-value thing on anyone’s list.

Where these crates differ

• Zero runtime dependencies in er7, enforced by a test. The dependency table is empty and a test fails if that changes. Criterion lives in a separate unpublished workspace member so that even the development dependencies stay empty.
• A command-line tool, not only a library. Every label in its output is a valid query you can paste back in.
• Redaction as a first-class, separately published thing. er7-redact has no equivalent in the list above.
• Specification-first development. Every behaviour is a numbered rule, and each rule names the test that enforces it — checkable by reading the tree.
• Five licenses, where the alternatives above are Apache-2.0 or MIT.

The honest comparison, because it is what most teams actually do and it is sometimes right.

When hand-rolling is right: you read one field, from one sender, whose messages you control, in a script that will not outlive the week.

When it stops being right, in roughly the order teams discover it:

1. Escape sequences. \F\ is a literal | in a value. Splitting on | cuts a value in half, silently.
2. The MSH segment is special. MSH-1 is the field separator, so the segment does not index like any other.
3. A sender who does not use the default delimiters. They are declared in MSH-2 and you are expected to read them, not assume them.
4. Repeating fields. PID-13 may be one phone number or four, and the code that assumed one is now wrong.
5. Components and subcomponents go two levels deeper.
6. Writing. Now you need to escape on the way out, and an unescaped delimiter breaks the message for everyone downstream.
7. Round-tripping. Someone asks why the message that came out is not byte-identical to the one that went in, and the answer is a week.

The pitch is not “your split is wrong”. It is that the seven items above are exactly what er7 implements, in about the same runtime cost, with a test for each and no dependency to audit.

Stated here so no comparison implies otherwise.

• No validation or conformance checking. Structure is parsed; content is not judged.
• No typed segment model. No generated class per segment.
• No transport. No MLLP, no TCP, no SOAP.
• No HL7® FHIR® standard support, and no v2-to-FHIR® mapping.
• No clinical claim. These are encoding libraries, not medical devices, and nobody has certified them.

Each crate’s own specification states its scope and its non-goals precisely; that is the authoritative version of this list. The full write-up, with the reasoning, is COMPARISONS.md in the repository.