AstrolexisAI β’ Sovereign AI: Kulvex self-hosted platform KCode local terminal coding agent. 100% GPU, no cloud. mnemo models. Privacy & control first.
Joined March 2026
- Tweets 63
- Following 17
- Followers 7
- Likes 22
2 Photos and videos
Pinned Tweet
Apr 25
Built a code auditor that bundles tree-sitter grammars for 8 languages directly into a single static binary. ~12 MB of grammar wasm, no install dance, runs entirely offline.
Open source: github.com/AstrolexisAI/KCodβ¦
1
1
2
203
May 21
The math on cloud LLMs flipped.
2024: cheaper to call someone's API.
2026: cheaper to run on your own hardware after 4-8 months.
I wrote up cost, privacy, latency, vendor lock-in β the case for running LLMs on-device in 2026.
astrolexis.space/blog/why-onβ¦
46
Apr 30
Official KCode cascade demo thread
πTwo real runs with v2.10.406:
β’ NASA openmct β 13 candidates, 0 false alerts
β’ OWASP Benchmark SQLi β F1 0.829, 100% recall, cascade fully engaged Grok-4-fast Opus-4.7 ensemble beating single-model Opus by a mile.
x.com/curly60e/status/204991β¦
Two cascade verifier runs on KCode v2.10.406. Same binary, same flags.
π°οΈ NASA Mission Control: 13 candidates β 0 false alerts
β οΈ OWASP Benchmark v1.2: 15 candidates β 5 confirmed via ensemble
The verifier knows the difference between clean code and real bugs.
π§΅
79
Apr 30
Official KCode thread on the SpaceX Cursor deal π
Cursor writes code at lightning speed.
KCode makes sure that code has no critical vulnerabilities. In real security analysis, KCode is already a step ahead.
SpaceX just struck a deal with Cursor β up to $60B option to acquire.Cursor is outstanding at writing code fast.But if you actually care about that code being secureβ¦ KCode is superior. Straight up
87
AstrolexisAI retweeted
After many hours of benchmarking and fighting with LLMs...
KCode v2.10.406 is finally out with proper cascade ensemble.
This is the one I've been waiting for.
Great work team @AstrolexisAI π₯
@hackerfantastic curious what you think about the Grok Opus combo.
Apr 30
π KCode v2.10.406 is now available.
We just shipped and fully verified end-to-end cascade ensemble (Grok-4-fast β Opus-4.7).
Real results on OWASP Benchmark SQLi (1,494 findings):
β F1: 0.829
β Precision: 69.8%
β Recall: 99.6%
β Cost: ~$17
kulvex.ai/kcode/benchmarks
1
3
8
794
Apr 30
π KCode v2.10.406 is now available.
We just shipped and fully verified end-to-end cascade ensemble (Grok-4-fast β Opus-4.7).
Real results on OWASP Benchmark SQLi (1,494 findings):
β F1: 0.829
β Precision: 69.8%
β Recall: 99.6%
β Cost: ~$17
kulvex.ai/kcode/benchmarks
1
4
897
Apr 30
Download v2.10.406 (Linux, macOS, Windows):
β kulvex.ai/kcode
Recommended command to try the ensemble:
kcode audit . -m grok-4-fast-non-reasoning --fallback-model claude-opus-4-7
1
3
94
Apr 30
This release is another step toward our goal:
Best recall in the industry, full control over noise, and reasonable costs.
Huge thanks to everyone testing and giving feedback.
What do you think of the new cascade? Would you use it in CI?
kulvex.ai/kcode/benchmarks
2
28
AstrolexisAI retweeted
You can try this. Totally auditable code. x.com/astrolexisai/status/20β¦
Apr 27
We ship KCode, a security audit tool.
Today we audited it against itself and found a bug we'd flag in someone else's code: shell injection in our own PR generator.
Fix shipped in v2.10.380. β
1
1
3
158
Apr 27
We ship KCode, a security audit tool.
Today we audited it against itself and found a bug we'd flag in someone else's code: shell injection in our own PR generator.
Fix shipped in v2.10.380. β
1
3
263
Apr 27
Update size from v2.10.379:
141 KB downloaded
0.8s end-to-end
exact SHA match
Full binary: 117 MB. Delta = 0.1% of size.
Verified E2E against the live manifest at kulvex.ai. No npm, no GitHub Releases β self-hosted auto-updater.
1
1
38
Apr 27
Update: kcode update
Install: curl -fsSL kulvex.ai/install.sh | sh
For tiny deltas: apt/brew/dnf install bsdiff
Repo: github.com/AstrolexisAI/KCodβ¦
Release: github.com/AstrolexisAI/KCodβ¦
1
22
AstrolexisAI retweeted
The most useful part of the internal audit rounds: found 4 real bugs in my own patterns before they shipped.
`new Function(p)` was silently missed (the query only matched the bare-call form). `\eval($p)` PHP namespace-prefixed too.
Tree-sitter gives you the AST. You still own the iteration.
Apr 25
Built a code auditor that bundles tree-sitter grammars for 8 languages directly into a single static binary. ~12 MB of grammar wasm, no install dance, runs entirely offline.
Open source: github.com/AstrolexisAI/KCodβ¦
1
2
70
Apr 25
Built a code auditor that bundles tree-sitter grammars for 8 languages directly into a single static binary. ~12 MB of grammar wasm, no install dance, runs entirely offline.
Open source: github.com/AstrolexisAI/KCodβ¦
1
1
2
203
Apr 25
Single static binary. No backend. No telemetry.
Source code never leaves your machine. The LLM call (when used) only sees the snippet being verified β not the surrounding repo.
Apache 2.0.
1
1
33
Apr 25
Looking for feedback on patterns I'm missing β especially from people who've done audit work professionally.
Pattern library lives at src/core/audit-engine/ast/ if you want to PR a new sink for your favorite language.
github.com/AstrolexisAI/KCodβ¦
1
2
35