Has anyone here
already built it?
Ask HomeGrown before you build. It tells you whether another team has already built it, who to ask, how ready it is and how to get the code. That works even for repositories you can't open.
$ git clone https://github.com/tancred27/home-grown
Has anyone built a Kafka load tester for k6?
From a test run on Grafana's public repositories. Public code has no org chart, so the teams, people, managers and access rules in that test are generated.
- 94%
- of answers useful on the first try, across 92 real-world questions
- 0
- times it told someone to reuse something that didn't fit
- 100%
- of known cases where two teams built the same thing, spotted
- 0.1 s
- to search 100,000 projects, showing each person only what they may see
How it works
It keeps a record of every component, and every search runs as the person asking.
- 01
Watches repos
Rereads only the folders a push changed.
- 02
Reads the code
Splits a big repo into the parts people reuse.
kind → CLI, cluster library, kindnetd, Go utilities - 03
Records it
Purpose, owner, stage. Every note links to its file or commit.
- 04
Searches, then checks
By keywords and by meaning. Then an AI model reads the best 20 matches and picks up to three, each with a reason. If your question is unclear, it asks back.
Which cloud is the storage account in? - 05
Shows its work
Why each answer came back: what matched, how it ranked, and the AI's reason.
homegrown explain <id>
Everyone can find it. Only some can open it.
A map of everything your company has built, by line of business. The database checks every row, so people learn a private project exists and who to ask, while its code stays shut. Admins can also turn on a view of where teams built the same thing twice.
| Detail | Another team | Its own team |
|---|---|---|
| What it does, who to ask | Visible | Visible |
| Facts marked sensitive | Hidden | Visible |
| Reading the code | Ask for access | Allowed |
Your coding agent checks first
Claude Code, Cursor and other MCP agents ask HomeGrown before writing new code. If the code is private to you, the agent tells you who to ask instead of writing a copy.
claude mcp add homegrown -e HOMEGROWN_USER=alice -- python -m homegrown.mcp_server
Add Kafka produce and consume steps to our k6 tests.
team-xk6-kafka already built this: xk6-kafka, in beta. Its code is private to you, so I won't write a new Kafka client yet. Shall I draft an access request to Engineer d94c3a, who committed to it on 5 November?
xk6-kafka already does this, and you can read it. I'll build k6 with it instead of writing a new Kafka client.
Tested on real open-source code
Every number links to where it was measured, weak spots included.
| What we measured | Result | Where to check |
|---|---|---|
| Answers useful on the first try, across 92 real-world questions | 94% | 70% good, 24% okay, 6% bad. It never recommended reusing something that didn't fit · results |
| The right project among the first five results, out of 5,000 | up to 92% | 85% on test questions, 92% when asked casually · results |
| Spotting two teams that built the same thing | 100% | all 20 known cases; every group checked by hand was right · results |
| From a code change to searchable, 95% of the time | < 4 min | replaying an hour of all public GitHub activity · results |
Not there yet
- No company runs it in production yet. There's no single sign-on and no permission sync.
- Who to ask needs an export of your employee directory; there's no automatic link to Okta or Entra yet.
- Accuracy drops as a company grows, mostly from look-alike projects.
- Reading source code and spotting repeated builds are optional, and haven't run on a real company's code yet.
Questions
Does our code leave our network?
Only if you choose a hosted model. Search runs on your machines. With the code reader on, short extracts (declarations and their comments, with keys, passwords and internal hostnames removed) go to the model you pick. With a local model through Ollama nothing leaves.
What does it cost?
Nothing: it's Apache-2.0. A hosted model adds two calls per question, and the code reader a cheap call per changed part of the code. Bring your own key (Anthropic, Gemini, Groq or Ollama), or run with no model at all.
Where can it read from?
GitHub repositories, including their source code if you turn that on, plus Confluence and Notion pages. GitLab and Bitbucket are planned.
Who does it tell me to ask?
The two people who committed to it most recently and are still in your company directory, with their manager. If no one fits, the owning team's lead.
Does it need our directory?
Only for who to ask. You import a snapshot: id, name, team, manager, line of business and whether someone is active. Emails and logins are kept only as hashes, and nothing else is stored.
How does it know what's in production?
From releases and from use: a library two other teams depend on counts. Until an owner confirms it, a stage is shown as inferred.
Could someone see what they shouldn't?
The database checks every query. Tests run every test user through real searches and fail if anything outside their access comes back.