Sovereign AI · made in India · zero dependence
Software that builds itself.
And never leaves India.
code.in is the sovereign AI platform that plans, writes, ships and attests your software — on frontier-class specialised models, on in-country compute, with zero foreign dependency. Your code. Your data. Your borders.
Why sovereign
Platform sovereignty in the agentic era
Most tools send your prompt abroad, run it on a foreign model, and deploy it abroad. code.in keeps every byte inside India's borders.
Your prompt exits the border
Data leaves the country the moment you hit send. You don't control the model, the compute, or where the build lands.
Stays in India, start to finish
Residency attested, build signed. The whole pipeline — model, compute, deploy — lives inside India.
The platform
Agentic engineering, 10× faster
There's more to engineering than writing code. code.in compresses delivery so your teams can do what AI can't — decide what to build next.
The evidence
Every build ships with proof
Not a promise. A signed, tamper-evident record you can hand to procurement.
| System | rail-booking |
| Environment | india, production |
| Data class | Confidential |
| Residency | in-country, attested |
| Access model | RBAC, 5 roles |
| Foreign API | none |
Attested on every build
Every deployment produces a signed residency record confirming your code never left India's borders.
Append-only, tamper-evident
Every action is logged in sequence. Nothing is editable. The trail is the proof.
RBAC defined at build time
Roles and permissions are mapped as the system is engineered, not bolted on after.
Benchmarks
Closing the gap to the frontier — and open labs
Same harness, same discipline, measured on the same benchmark the frontier labs use. Every code.in number below is a real scored run of our own sovereign harness — a different model per track, no third-party scaffold.
| Model | Lane | Score |
|---|---|---|
| Claude Opus 5 | Closed frontier | 96.0% |
| Claude Opus 4.8 | Closed frontier | 88.6% |
| DeepSeek V4 Pro | Open weight | 80.6% |
| Qwen3.7 Max | Open weight | 80.4% |
| MiniMax M2.5 (own harness) | Open weight | 80.2% |
| code.in + MiniMax M2.5OURS | Open weight | 70.0% |
Our sovereign harness drove MiniMax M2.5 to 70.0% (350/500) on SWE-bench Verified — trailing the same model's 80.2% on its own harness, with the open-weight frontier at ~80 and the closed frontier at ~88–96. We're closing the gap with a model-agnostic harness and smart tokenomics: swap in any open-weight model, pay only for what you resolve, and run it all in-country.
| Model | Lane | Score |
|---|---|---|
| code.in + GLM 5.3 FlashOURS | Open weight | 63.0% |
| Claude Opus 4.6 | Closed frontier | 62.7% |
| MiniMax M2.5 (native) | Open weight | 56.3% |
| GPT-5 | Closed frontier | 54.3% |
On Lite, our harness with GLM 5.3 Flash scores 63.0% (189/300) — leading the table, ahead of MiniMax M2.5's native 56.3% and Claude Opus 4.6's 62.7%. We're already ahead on the smaller split.
Two tracks tell one story. On SWE-bench Lite we already lead — 63.0% with GLM 5.3 Flash. On SWE-bench Verified we trail the frontier by ~10–25 points with the same harness. Closing that gap is scaffold work: reproduce-test-first, regression filtering, best-of-N, hierarchical localization. We port the patterns, not the code, and run them all inside India.
Who it's for