From commit to credit, in four passes.
Three specialized agents, one chronological backbone, one audit-ready claim.
Pulls diffs from Git, tickets from Jira, wages from payroll. CodeClaim gives us diff-only evidence — no source exposure.
Applies the IRS §41 four-part test to every project: technological in nature, eliminating uncertainty, process of experimentation, permitted purpose.
Sums qualified wages, supplies, and contract research across the project timeline. Flags exclusions (funded research, foreign research, post-production).
Drafts the IRS-ready narrative using Microsoft RD-Agent chronology and your actual commits. Big-4 reviewed templates.
Every project is scored on all four prongs.
If any prong fails, the project is excluded — and we tell you exactly why.
The research must rely on principles of the hard sciences — engineering, computer science, physics, biology. We map commits to STEM disciplines and reject pure marketing or business-method work.
Did the team face uncertainty about capability, methodology, or appropriate design at the outset? We surface design docs, spike branches, and uncertainty resolution in PRs.
Iterative evaluation of alternatives — A/B tests, prototypes, model comparisons, benchmarking. We trace hypothesis → test → outcome from your commit history.
New or improved business component — product, process, technique, formula, or software. We tie each project to a tangible deliverable and reject pure internal tooling.
Estimate your credit in 60 seconds.
Plug in headcount, engineering payroll, and contractor spend. R&DCredit AI applies the regular credit (6% / 10% AMT offset) vs. the Alternative Simplified Credit (14% / 6% AMT offset) and picks the bigger number.
Illustrative only. Final credit determined by your CPA. R&DCredit AI is not a tax preparer.
Statutory checks before a single number leaves the pipeline.
Three guarantees separate a defensible claim from a guess: live IRS lookup, deterministic tax math, and a filled Form 6765.
Before filing, Agent 4 queries the IRS Taxpayer MCP server for the current Revenue Procedures, §41 ceilings, and any carry-back/carry-forward rules. If a number falls outside live IRS guidance, the claim is flagged before it leaves the building.
QRE metrics are parsed into corbt/2025-tax-engine — a rules-based Python tax engine. No LLM hallucinates your credit: every line is computed deterministically, audit-defensible, and reproducible bit-for-bit.
The final step maps computed numbers directly into an open-source PDF form filler to output a completed IRS Form 6765 (Credit for Increasing Research Activities) — ready for e-file alongside your technical narrative.
Guardrails programmed into every agent.
Devastating audit failures start with two things: forgetting geography, and double-dipping credits. R&DCredit AI rejects both by default.
The ingestion agent automatically filters out any contractor or employee wages for work performed outside the United States. Foreign research is strictly ineligible under IRC §41 — we strip it before it reaches Agent 2.
If you claimed the Employee Retention Credit (ERC) for any overlapping period, Agent 2 explicitly checks and subtracts those wages to avoid violating IRC §280C(c). You can't claim the same dollar twice.
File your strongest R&D claim yet.
Connect GitHub, Jira, and payroll. R&DCredit AI runs the four-part test, sums your QREs, and drafts the narrative — usually in under 48 hours.