Inside Fullscript's software factory
19% of January PRs created by Nitro, the in-house agent that lets PMs ship UX fixes and caught bugs in past incidents when run retroactively.
Fullscript at a glance
The company
| Sector | Digital health platform for practitioners and patients (supplements, wellness) |
|---|---|
| Employees | Not disclosed as a discrete count |
| Customers | 100,000+ practitioners 1 |
| Revenue | Not disclosed |
Fullscript is a private company; headcount and revenue are not publicly disclosed.
The engineering organisation
| Engineering size | 180+ engineers 1 |
|---|---|
| AI tooling since | Mid-2025, when Nitro development began; late October 2025 official launch 1 |
| Tool policy | Nitro built in-house; Claude Code as swappable backbone 1 |
| Agents in the SDLC | Background coding agent; automated code review; issue implementation 1 |
| Reported adoption | 19% of January 2026 PRs created by Nitro; 45% of PRs voluntarily tagged for Nitro review 1 |
| Published metrics | PR creation share, review adoption rate 1 |
The 19% figure measures PRs created by Nitro, not lines of code or autonomous authorship. Human review and CI remain required on every change 1.
People
What they set out to solve
Fullscript runs a mature Rails monolith with a React frontend. It's over a decade old, with millions of lines of code and 180+ engineers contributing regularly. That scale and history is an asset, but it also meant off-the-shelf AI tools didn't quite fit 1.
The specific barriers were a self-hosted GitLab instance where many tools are GitHub-first and integration was unavailable or limited, monolith complexity where the codebase needed execution in the right environment with the right access and drop-in solutions couldn't easily navigate the setup, a workflow mismatch where Fullscript uses Linear for issues and GitLab for code while many tools assume one platform or the other, and internal infrastructure where they wanted something that could access internal APIs, run against staging environments, and use their secrets management 1.
Leadership gave a small team a mandate: figure out if this could work. It was a gamble, pulling developers off product work to chase an idea that might not pan out. Within six months, they had something real 1.
Maturity level
Assigned by testing published facts against fixed criteria, and capped by what the sources state.
How Fullscript's software factory works
Nitro is an agentic developer in the cloud integrated with all of Fullscript's tools. It doesn't live on laptops. It lives in the issue tracker and code review system. Assign it an issue, and it opens a pull request. Add it as a reviewer, and it leaves comments 1.
How engineers use it
- In GitLab: Mention @nitro in any issue or PR comment.
- In Linear: Assign Nitro to an issue or tag it in comments.
No new tool to learn. No context switching. Just work where engineers already work 1.
What it can do
Code reviews. Fullscript requires two reviewers on every PR with at least one approval to merge. That's a lot of review work across 180+ engineers. Nitro helps carry that load. Add Nitro as a reviewer and it leaves inline feedback on conventions, performance issues, and security concerns, consistent every time. Developers are voluntarily tagging Nitro on 45% of PRs 1.
Catching what humans miss. Fullscript retroactively ran Nitro Review on past pull requests that had caused production issues. Nitro identified the bugs. If they had been using Nitro on those changes, they could have prevented real problems. Not because humans are bad at reviews, but because AI doesn't get tired, doesn't skim, doesn't rush before a meeting. It applies the same attention at 11pm Friday as 9am Monday 1.
Issue implementation. Assign a Linear issue to Nitro. It reads the issue, writes code, and opens a PR. Works best for well-defined, scoped tasks 1.
Answering questions and writing tickets. Ask Nitro to dig into an issue and it'll explore the code, understand the context, and respond in the thread. From there, have it update the issue with better context or split a vague issue into smaller, well-spec'd tickets in Linear with proper details and code references 1.
Expanding who can contribute. A PM describes a visual bug or small UX issue. Nitro opens a PR ready for review. A designer requests a copy change. Working code, ready for review. An operations lead fixes an unintuitive aspect of an admin tool that's been bugging them for months 1.
Addressing feedback without context switching. Put up a PR, move on to the next task, let feedback accumulate. When ready to address it, leave a comment, "@nitro address the feedback," and it pushes the fixes with full context on the discussion. No branch switching, no context thrashing 1.
Under the hood
Each request spins up a fresh environment with a clean clone and no state leakage between runs. Claude Code powers Nitro as the backbone, though this is swappable for Codex, Gemini, OpenCode, or others; Claude has been the best fit so far. GitLab and Linear events trigger Nitro automatically via webhooks, and responses post back to the source. Fullscript already had facilities for developers to stand up testing environments on the fly, and they gave that ability to Nitro too 1.
What works and what doesn't
The quality bar doesn't change: human reviewer on every PR, developer approval to merge, CI and E2E tests have to pass. Nitro's output goes through the same process as anyone else's code 1.
Where Nitro shines: well-defined, scoped tasks. Fixing flaky specs, small UX changes, resolving N+1 queries, bug fixes with stack traces, feature flag cleanup, scoped tasks within larger features, prototyping from designs 1.
Where it struggles: ambiguous tasks (vague requirements produce vague output), deep architectural work (doesn't understand system-wide tradeoffs well), large context needs (changes touching many files can exceed useful context) 1.
Operating Fullscript's software factory: teams, roles, and skills
Nova is the shared skills library that teaches coding agents Fullscript's standards. It's automatically added for users in their environment. EKG is the internally developed MCP gateway with role-based access and PII/PHI redaction built in 5.
The pattern that emerged from large refactoring work: directors started assigning Linear tickets directly to Nitro. Instead of treating it like a second-class tool, teams started treating it like a real engineer: Nitro picks up the ticket, performs the migration, engineers review and test. The bottleneck stopped being "who has time to migrate 50+ modals" and became "how quickly can we perform code reviews and test the code" 2.
This shifts engineers higher up the stack, toward setup, review, and higher-value work. The valuable work is no longer manually updating the 50th modal implementation. It becomes designing better migration strategies, defining architectural patterns, writing high-quality skills, reviewing complex edge cases, and testing the implementation 2.
Implementation journey
- Mid-2025Nitro development begins. Leadership gives a small team a mandate to figure out if building an in-house agent could work 1.
- Late October 2025Nitro officially launches 1.
- January 202619% of PRs created by Nitro; 45% of PRs voluntarily tagged for Nitro review 1.
- February 2026Post published. CEO has shipped three changes via Nitro, including a bug fix and a new mobile app feature 1.
Measuring Fullscript's software factory
The headline figures are PR creation share and review adoption: 19% of January 2026 PRs created by Nitro, and 45% of PRs voluntarily tagged for Nitro review 1.
No stated percentage of AI-authored code appears in the sources. The 19% figure measures PRs created, not lines or autonomous authorship. Human review and CI remain required on every change 1.
Lessons learned
Early on, Nitro was used to fix some N+1 queries. It didn't have a full understanding of some nuances in the setup, how certain patterns could surface N+1s in ways that weren't obvious. The fix it proposed looked reasonable, passed human review, and shipped. Then it caused 500s in production. They reverted it 1.
That failure clarified something: Nitro can only be as good as the context it's given. Conventions that lived in people's heads? Nitro didn't know them. And honestly, neither did some developers at the time 1.
To make Nitro effective, they had to document things they'd never written down: conventions, patterns, the nuances that caused that N+1 incident. This documentation helps new developers just as much as it helps Nitro. Teaching an agent exposed gaps in how they taught humans. They are writing more documentation than ever before, partly due to time saved by these tools and the incentive to make them better 1.
What's next for Fullscript's software factory
Three directions are described: faster and more responsive by reducing environment spin-up time and less waiting, mid-task interaction where the current fire-and-forget model becomes back-and-forth with follow-up, course correction, and iteration, and more platform integrations including Slack triggers and notifications, a CLI, and a front end for more ways to trigger and interact with Nitro. The goal is more like a teammate you can ping, less like a batch job you submit 1.


