Inside Alan's software factory
How Alan's Hopper authors 25% of merged PRs and lets 80% of employees ship code, from product managers to analysts, through one MCP-gated agent.
Alan at a glance
The company
| Sector | "Where insurance, care, and prevention finally work together" 4 |
|---|---|
| Employees | Not disclosed as a discrete count |
| Customers | 1 million+ members in France, Belgium, Spain, and Canada 4 |
| Revenue | Not disclosed |
Alan is a private company; employee count and revenue are not publicly disclosed.
The engineering organisation
| Engineering size | Not disclosed as a discrete count |
|---|---|
| AI tooling since | Spring 2026, when Hopper launched as a codebase Q&A tool 1 |
| Tool policy | Hopper built in-house; local agents like Claude Code and Codex also approved through MCP Gateway 1 |
| Agents in the SDLC | Cloud coding agent; shared sessions; sites and dashboards for prototyping and analytics 1 |
| Reported adoption | 80% of employees use Hopper; 25% of merged PRs from Hopper 1,2 |
| Published metrics | PR-authorship share, employee adoption rate, sites created 1,2 |
The 80% adoption figure includes non-engineers: product managers, designers, analysts, and operations staff, not just the engineering organisation 1.
People
What they set out to solve
Hopper started as a codebase Q&A tool in spring 2026. The stated goal was letting engineers, product managers, and designers ask questions about how Alan's code worked and plan a change before writing it, without interrupting an engineer. The most frequent question users asked was "When can it build?" 1.
The summer release of Build Mode addressed that directly: Hopper could start from a Slack message or a Linear issue, make the change in its own environment, run tests, and open a pull request. Within two weeks, 13% of commits were co-authored by Hopper, and 35% of engineers had merged work started in Hopper 1.
Then users pushed it past what the team had designed for. Product managers prototyped flows. Designers shared interactive mockups. Security engineers published reports. Analysts asked questions about data and wanted answers to stay up to date. The decision that followed shaped everything else: keep a single agent rather than building one per job 1.
Maturity level
Assigned by testing published facts against fixed criteria, and capped by what the sources state.
How Alan's software factory works
At the core sits Hopper, a cloud coding agent reachable from a web app, Slack, Linear, or GitHub. Sessions run in the cloud, keep going when you close your laptop, and can be continued by colleagues. A product manager can clarify a requirement in a session an engineer started, and the engineer can pick it back up after review 1.
The stated design philosophy is one agent, not one per job. What changes from one job to another is mostly context and tools. A data question needs metric definitions and warehouse access. A product question needs research notes and customer feedback. A code change needs the repository and a shell. So Hopper is one agent loop plus tools, skills, and connectors 1.
The MCP Gateway
Every tool call from Hopper, and from approved local agents like Claude Code or Codex, goes through one Alan endpoint: the MCP Gateway. It decides which tools a session can see, whether a call is allowed, needs confirmation, or is denied. If the call is allowed, it decides which credential to use, an organization connection shared by everyone or the user's own account. Then it records the result 1.
One agent with every connector carries a cost: hundreds of tool definitions would fill the context and make the model worse at picking the right one. So the agent does not see them up front. It gets two tools, one to search the catalog and one to call what it found, and the search only returns tools the session is allowed to use. The price is an extra round trip before the first call to a new tool 1.
The Gateway is what made the single-agent approach practical.
Tim Petricola
What Hopper produces
Some outputs land in the tools Alaners already use: a pull request, a Linear issue, a Notion page, a Slack message, an email, or a calendar event. Some live in Hopper itself 1.
Sites are small web pages that Hopper builds and hosts. They get a URL, version history, and comments anchored to specific elements, so feedback points at the exact button or paragraph it is about. Hopper can then pick up those comments and iterate. Since sites shipped in the summer, almost 400 Alaners have created more than 1,700 of them, 1,200 in the last 30 days alone. Many teams now build a site in place of where they used to write a spec 1.
Dashboards are for numbers to follow over time. Hopper looks up the business definition in Alan's data ontology, writes the SQL, and draws the chart. If you want to keep following the number, you save it as a dashboard. The SQL is visible and editable, so anyone can check what a chart actually measures. Refreshing a dashboard does not involve a model: the saved queries simply run again 1.
Shared by default, safe by construction
Sessions, sites, and dashboards raise the same question: who can see them? At Alan, the default answer is everyone. The company operates on the principle of radical transparency: work is visible by default, so anyone can understand a decision and reuse what led to it 1.
Three decisions shaped permissions. One permission model covers sessions, sites, and dashboards, with organization-level settings (restricted, any Alaner can read, any Alaner can write) and per-person settings (view, write, manage). The agent acts as the person doing the work: if someone continues another person's session, the Gateway resolves the actor from the run in progress, never from the session's original owner. Credentials and approvals stay on the server: connector credentials never enter the agent's environment, and when a call needs confirmation, the Gateway stores the proposed tool and its arguments until a human approves 1.
Health data raises the bar further. Most sessions never touch any, but the entire execution path, from model inference to temporary files, is designed toward HDS, the French certification for hosting health data. The requirement already constrains which models Hopper can use: only AWS Bedrock and Azure deployments that meet European processing and retention requirements 1.
Architecture evolution
In October 2026, the team moved the agent out of the sandbox. The agent loop now runs in Alan's backend, and a sandbox starts only when a task needs a filesystem, a shell, or a browser. This change halved the median time before the agent starts responding 1.
Operating Alan's software factory: teams, roles, and skills
No team structure is disclosed. The engineering detail comes from one post by one author. Adoption is framed as pull rather than mandate: each step came from users pushing Hopper past what the team had designed for 1.
Skills are the mechanism for sharing workflows. They are versioned instructions for repeatable tasks, stored in Git or in an internal registry and maintained by the teams that know the work 1.
Implementation journey
- Spring 2026Hopper launches as a codebase Q&A tool 1.
- Summer 2026Build Mode released. Within two weeks, 13% of commits co-authored by Hopper and 35% of engineers merging Hopper-started work 1.
- Q1 2026Alan reports a 20–30% increase in PRs per engineer and pull requests merged by non-engineers growing 3.
- April 2026Operations AI agents platform published, with 15 tools and human-in-the-loop enforcement at the application layer 5.
- October 2026
Measuring Alan's software factory
The most consistently cited figure is PR-authorship share: 25% of merged pull requests come from Hopper as of October 2026 1,2. That figure measures merged pull requests specifically and is self-reported.
Adoption is measured by employee count: 80% of Alaners use Hopper. This figure explicitly includes non-engineers, product managers, designers, analysts, and operations staff who now ship alongside engineers 1.
A separate metric tracks site creation: almost 400 Alaners have created more than 1,700 sites, 1,200 in the last 30 days 1.
Lessons learned
The post frames value measurement as an open problem. PR share was a good first North Star for engineering, but it says nothing about a dashboard that changed a decision or a prototype that killed a bad idea early 1.
Sharing by default made permissions a design problem from day one. Each new capability, sites, dashboards, credentials, required the same policy decisions before launch, not after 1.
The one-agent-not-one-per-job decision carried its own risk: tool catalog size. Hundreds of tool definitions would degrade model performance, which led to the two-meta-tool pattern (search, then call) that adds a round trip but keeps context clean 1.
What's next for Alan's software factory
The post lists four open problems 1:
- How to measure value beyond PR share.
- How to organize more than 1,000 new sites per month.
- How far sharing should go: some sites need to be shared outside Alan, some dashboards should only show what the viewer's own data access allows.
- How to open up health-data use cases safely.
The operations platform post describes a north star of 80% automation, with the operations team iterating on agents themselves using the platform, and engineering's job becoming enhancing the platform rather than iterating on individual agents 5.


