For Canadian teams documenting SR&ED work

Build the SR&ED record while the work is happening.

T661.ai turns the history already in your repositories into monthly technical logs, annual narrative drafts, a checked expenditure worksheet, and a package your accountant can read.

Start your record See a finished sample

From $49 a month, never a share of your claim. The first completed month is free and needs no card. The 14-day trial starts only once that preview has found evidence and built a log.

Read-only GitHub App
No source files or diffs stored
Your team confirms every final record
Cancel any time

The work happened. The record is scattered.

Three days from one month at a fictional company. Two are already written down. The third is the one a reviewer asks about first.

14 March 2026 · A commit message

Reverted the clause joiner; prosody artifacts at the join. Keeping the harness.

The failed approach, written down at the moment it failed. Nobody will remember this in February.

22 March 2026 · A pull request

Benchmarks first-audio latency against the sentence-level baseline across two voices.

The method and the measurement, already described by the person who ran it.

31 March 2026 · Nowhere

Why any of it was uncertain at the start of the month.

The one thing the repository cannot know, and the one thing a reviewer asks about first.

One source today

GitHub, through a read-only App you install on the repositories you choose. The rest are planned, and saying so is more useful than a row of logos.

GitHub · available
Self-hosted Git · planned
GitLab · planned
Bitbucket · planned

Every sentence says where it came from

This is a real working paper over a fictional company. Resolve the marker to see what review actually looks like.

northstar-voice__2026-03.md

Fictional company, real structure

Work performed — March 2026

Every statement carries where it came from. That is what a reviewer follows.

Observed · COMMIT 4f2a1c

Split the text-to-speech stream at clause boundaries instead of sentence boundaries, and measured first-audio latency against the sentence baseline.

Observed · COMMIT f81d3a

Reverted the clause joiner after prosody artifacts appeared at the join, keeping the measurement harness in place.

Inferred · MONTH 2026-03

The month's work addressed an uncertainty about whether perceived latency could be reduced below 800 ms without degrading prosody.

Needs review · VERIFY · unsupported figure

First-audio latency at the 95th percentile fell from 1.9 s to 740 ms across all supported voices.

The benchmark in the evidence covers two voices, not all of them. A person decides how to say it.

One statement still needs a person.

ObservedInferredNeeds review

Observed came out of the repository. Inferred was written from what was observed. Needs review is waiting on someone at the company. Nothing is filed until every marker is gone.

What your accountant opens

One locked snapshot, organized the way a reviewer reads it. Working papers, not a filed return and not a tax opinion.

Annual narrative drafts

For T661 Part 2, with the word limits and the continuity checks between years.

Monthly technical logs

Each month as it was reviewed and confirmed, with the markers resolved.

Expenditure worksheet

Salaries, project allocation, available time, and the rates, calculated rather than drafted.

The underlying evidence

Commit identifiers, dates, authors, and the pull requests behind every cited statement.

Supporting documents

Whatever the company uploaded, kept beside the claim it belongs to.

Revisions and a file list

Who changed what and when, with hashes that show if a file moved afterwards.

The expenditure worksheet is calculated, not drafted. One module owns every arithmetic relationship on the form and the screens print it, so the numbers are the same every time the package is built.

A quiet month is an answer, not a gap

A month that was reviewed and found no R&D work is finished. A month nobody collected is not. At year end those look identical in a folder, so the product never lets them look alike.

Nov
2 projects
Reviewed and confirmed
Dec
1 project
Reviewed and confirmed
Jan
2 projects
Reviewed and confirmed
Feb
No R&D work found
Reviewed, nothing to claim
Mar
3 projects
Reviewed and confirmed
Apr
Not collected
Needs collecting

Flat monthly pricing

The software costs the same whatever your claim is worth. Both plans include the same workflow; the difference is how many repositories you can connect at once.

Solo

$49 a month

Up to 3 repositories, unlimited members.

Team

$99 a month

Up to 10 repositories, unlimited members.

Compare the plans

Questions

What does the CRA actually require?

The CRA looks at two things: why the work was needed, and how the investigation was carried out. It also says the strongest supporting records are dated and specific to the work. T661.ai preserves the original dates from Git history and labels any later AI-written draft. GitHub evidence can support your account of the work; it does not decide eligibility. Your company still provides the technical context, reviews every statement, and confirms hours from its own records.

Does T661.ai read my source code?

The GitHub App needs read access because that is the permission GitHub uses for commit information. T661.ai stores evidence about the work: commit identifiers, messages, authors, dates, branches, file paths, change counts, and relevant pull-request text. It does not store source files or diffs. If GitHub includes a code fragment in a response, T661.ai discards it immediately and never sends it to Claude.

Is the AI going to invent results?

Claude writes a draft; it does not decide what is true or whether work is eligible. Every cited commit identifier is checked, and a ⚠ VERIFY marker goes on any statement the evidence does not support. You correct or remove every marker, enter hours from company records, and make the final confirmation yourself. The figures on the expenditure worksheet are calculated in code, not written by the model.

Is this tax advice? Will it guarantee my claim?

No, and no. T661.ai prepares technical working papers from your repository history. Eligibility determinations, expenditures, and filing remain with you and your SR&ED advisor or accountant. The product preserves dated evidence, exposes unsupported statements, and blocks an incomplete package inside its supported scope.

Which claims can be finalized in the product today?

A British Columbia Canadian-controlled private corporation using the proxy method, where overhead is calculated from eligible salary, and where the claim does not involve associated corporations, materials, capital property, late-paid amounts, work performed outside Canada, or work performed for another party. Every other claim can still build and export its evidence for an accountant to complete. “Ready for accountant review” means the files are organized and the required checks have passed. It does not promise that an accountant or the CRA will accept the claim unchanged.

What if my repositories are not on GitHub?

GitHub is the only source supported today, through a read-only GitHub App. Self-hosted Git, GitLab, and Bitbucket are planned. Tell us what you use so we can order the work.

Can I go back and cover earlier months?

Yes. A log can be built for any completed month your Git history covers. The original commit dates stay intact, and the later summary records separately when it was generated and reviewed. After setup, a new draft is scheduled when each month ends, and you are told if it cannot finish.

Why keep using it after the first draft?

A month is easier to review than a year reconstructed from memory. New work is captured while the context is fresh, failed approaches stay visible, deadlines stay in view, and every revision and supporting document stays attached to the claim year it belongs to.

Start with one month

Connect one repository and let it build a completed month. If there is nothing worth claiming in it, you will have found that out in an afternoon rather than in February.

Start your record