Meta Launches Internal AI Coding Tool MetaCode to Rival Anthropic and OpenAI
0xBroomberg
Meta is requiring thousands of engineers to submit at least one code fix per week to train MetaCode, its in-house AI coding agent, aiming to close the gap with Anthropic and OpenAI while cutting billions in external tool spending.
What is MetaCode, and why is Meta building it?
MetaCode is Meta's in-house AI coding agent — an AI assistant that can write and revise code autonomously. It currently has about 7,000 weekly active users, all internal engineers.
This means → Meta wants to stop paying Anthropic and OpenAI for their tools. The company has already spent billions of dollars this year on external coding assistants like Claude Code and Codex.
In plain terms = rather than pay billions a year for someone else's coding AI, Meta would rather build its own — and keep the data in-house.
What role do engineers play in this?
Maher Saba, Meta's VP of applied AI engineering, sent an internal memo requiring every engineer to submit at least one code "diff" (a code-change record) to MetaCode per week.
These fixes are not just "user feedback." They feed directly into training the next-generation model, internally code-named "Watermelon."
This reflects a deliberate design: Meta is turning its engineers into data annotators. The people writing code are simultaneously feeding training data back into the AI — a closed human-model loop.
Is it working?
By the time the memo went out, engineer-submitted fixes had produced over 800 system improvements.
The latest model, Muse Spark 1.1, saw measurable coding-ability gains, with rising scores on DeepSWE — a benchmark that tests how well AI handles real software-engineering tasks.
To drive participation, Meta added gamification: engineers earn color-coded badges on their internal homepage based on how many fixes they submit. This means → Meta is not just encouraging people to "try it out" — it is turning data contribution into a quantifiable, competitive daily task.
How does this differ from the usual tech-company "dogfooding"?
Dogfooding — using your own product internally — is standard at Google, Microsoft, and OpenAI.
Meta goes further. It is not just asking engineers to "use the tool." It is systematically harvesting correction data and piping it directly into model training.
In plain terms = other companies say "try our new product." Meta says "use it, and while you do, help the AI do its homework."
Will MetaCode go external?
Mark Zuckerberg hinted on the Q2 earnings call that MetaCode may eventually be released externally, adding that "there are more coding and productivity tools on the roadmap."
This means → if MetaCode matures enough for external release, it will compete head-to-head with Claude Code and Codex — potentially turning a two-player AI coding-assistant race into a three-way contest.
The key test: whether a tool trained on internal-engineer data can adapt to the wildly diverse codebases and development environments outside Meta's walls.
Content is for reference only, not financial advice.