The R&D tax credit for devtools startups
The short answer
Devtools engineering, compilers, language servers, and build systems, qualifies for the R&D tax credit at a higher rate than most software categories, since deep technical uncertainty is the normal day job.
What qualifies, and what fights you
Devtools companies build the kind of software the R&D credit was written for. Parsers, compilers, language servers, and build systems require solving problems that do not have a known-good answer, which is the core of technological uncertainty.
The sticking point in devtools is usually smaller than in other industries: less of the work needs to be excluded. Documentation, routine dependency updates to the tool's own codebase, and simple bug fixes still do not count, but they tend to be a smaller share of the total engineering time than in a typical product company.
Open source maintainers who also run a commercial devtools company should split the two. Time spent triaging community issues or reviewing routine pull requests is not research. Time spent designing a new incremental compilation algorithm is, whether or not the resulting code is open source.
The four-part test, applied to devtools startups
Qualified purpose is clear: the work builds or improves the tool. Technological nature is rarely in question either, since compilers and language servers are applied computer science by definition.
The interesting part is elimination of uncertainty and experimentation. Building a language server means figuring out how to keep symbol resolution correct and fast across a codebase with millions of lines, while the developer keeps typing. There is no textbook answer for a given codebase's shape. Teams test different indexing strategies, measure latency, and iterate, which is exactly the process of experimentation the test asks for.
New to the test itself? Read what software work qualifies as R&D first.
Work that usually qualifies
Language server development
Building incremental parsing and symbol resolution that stays accurate and fast as a developer edits a large codebase in real time involves deep, unresolved technical problems.
Incremental compiler or build system
Designing a caching model that correctly invalidates only what changed, without over-building or under-building, requires solving correctness problems with no standard solution.
Static analysis engine
Building a linter or analyzer that tracks control flow and data flow across a codebase to catch a new class of bugs is applied research, not routine coding.
Dependency resolution algorithm
Building a package manager's version resolution logic, especially handling conflicting version constraints, is a constraint satisfaction problem with real technical uncertainty.
Large-scale indexing performance work
Redesigning how a tool indexes a monorepo with millions of files, so search and navigation stay fast, qualifies as performance engineering under uncertainty.
Work that usually does not
Writing user documentation
Documenting existing CLI commands and flags does not involve technological uncertainty, regardless of how much time it takes.
Routine dependency bumps in the tool's own build
Updating the versions of libraries the tool itself depends on, with no design change, is maintenance, not research.
Which expenses count
Wages for the engineers building the core tool, compiler, or language server count, prorated to time on qualifying design and implementation work. This usually includes most of a devtools engineering team, since the product itself is the research.
US-based contractors count at 65% of what you pay them. A contract compiler engineer brought in for a specific optimization project counts under the same rule as an employee.
Cloud compute used for CI infrastructure, build testing at scale, and performance benchmarking counts. Compute spent hosting the finished tool for paying customers in production generally does not.
A worked example
Hypothetical example. A devtools startup has 4 engineers earning a blended average of $170,000, spending about 80% of their time on core product research, since the tool itself is the research.
At 6 to 10% of total QRE, the federal credit lands between about $35,900 and $59,800. Under $5 million in revenue, the company can apply up to $500,000 of that against payroll taxes each year.