Run the same product inside your perimeter. Same engine, same retrieval, same MCP. Plug in your own LLM and your own Git server.
If your code can't leave the network (for regulators, IP, or your own security team), the cloud product is a non-starter. This is the same product running on your hardware.
Nothing leaves your VPC. Ever.
Source stays where you control the keys
No new vendors, no exception requests.
Self-hosted ships the same binary as cloud. The matrix below is exhaustive.
Full Context Engine
Multi-Repository Intelligence
Code Review Agent
Deep Research
MCP Server
API Access
GraphRAG Knowledge Graph
Incremental Indexing
Team Collaboration
RBAC & Permissions
Tenant Isolation (per-org KEK binding)
Encryption at Rest (AES-256-GCM envelope)
Custom LLM Support
SSO / SAML
Data Residency Control
Two install paths: one for getting started, one for production at scale.
Single-node deployment for smaller teams.
Up in under an hour on one node.
Horizontally scalable, high-availability deployment.
HA, autoscaling, thousands of repos.
All deployments include automated backup, monitoring endpoints, and health checks. Container images are pulled from our private registry (or your mirror).
Retrieval does the work, so a 70B-class open-weights model handles most queries. No H100 cluster required.
Most of what makes a frontier model look smart on code is context. We supply the context, you bring a smaller model that fits your GPU budget.
Tested with these. Anything OpenAI-compatible also works.
| Model | Notes |
|---|---|
| gpt-oss-120b | OpenAI open-weights generalist; strong default for coding/reasoning |
| GLM 5.1 | Zhipu AI flagship; long context, strong code understanding |
| Kimi 2.6 | Moonshot agentic model; large context window |
| DeepSeek V4 | DeepSeek MoE flagship; strong on coding tasks |
| Qwen 3 Coder | Alibaba long-context coder model |
gpt-oss-120b
OpenAI open-weights generalist; strong default for coding/reasoning
GLM 5.1
Zhipu AI flagship; long context, strong code understanding
Kimi 2.6
Moonshot agentic model; large context window
DeepSeek V4
DeepSeek MoE flagship; strong on coding tasks
Qwen 3 Coder
Alibaba long-context coder model
Hardware footprint depends on the chosen model and quantization. Pick a model that fits the GPU budget you already have. CodeAlive's retrieval lets a smaller model behave like a larger one with better context.
For hybrid deployments.
CodeAlive talks to any OpenAI-compatible API. That covers vLLM, Ollama, TGI, LocalAI, SGLang, and most internal LLM gateways. Point the LLM endpoint at your inference stack, drop in an API key, done.
Smaller model with good retrieval beats a frontier model without it.
No SOC 2 badge yet. Instead, here's what's actually in the codebase today.
Every component, where it sits in your network, and what it talks to.
┌─────────────────────────────────────────────────────────────────────────┐ │ YOUR SECURE PERIMETER │ │ ┌───────────────────────────────────────────────────────────────────┐ │ │ │ CodeAlive Platform │ │ │ │ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │ │ │ │ │ Web UI │ │ MCP Server │ │ REST API │ │ │ │ │ └──────┬──────┘ └──────┬──────┘ └──────┬──────┘ │ │ │ │ │ │ │ │ │ │ │ └────────────────┼────────────────┘ │ │ │ │ ▼ │ │ │ │ ┌─────────────────────────────────────────────────────────────┐ │ │ │ │ │ Context Engine │ │ │ │ │ │ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │ │ │ │ │ │ │ Indexer │ │ GraphRAG │ │ Query │ │ │ │ │ │ │ │ Service │ │ Engine │ │ Processor │ │ │ │ │ │ │ └─────────────┘ └─────────────┘ └─────────────┘ │ │ │ │ │ └─────────────────────────────────────────────────────────────┘ │ │ │ │ │ │ │ │ │ │ ▼ ▼ │ │ │ │ ┌─────────────┐ ┌─────────────────┐ │ │ │ │ │ Vector DB │ │ Your LLM │ │ │ │ │ │ (Embedded) │ │ (vLLM/Ollama) │ │ │ │ │ └─────────────┘ └─────────────────┘ │ │ │ └───────────────────────────────────────────────────────────────────┘ │ │ │ │ │ ▼ │ │ ┌──────────────────┐ │ │ │ Your Git Server │ (GitLab, Bitbucket, Gitea, etc.) │ │ │ (Internal) │ │ │ └──────────────────┘ │ │ │ │ ═══════════════════════════════════════════════════════════════════ │ │ NO EXTERNAL NETWORK TRAFFIC │ └─────────────────────────────────────────────────────────────────────────┘
CodeAlive itself, excluding any local LLM you run alongside.
Minimum (POC)
Production (small/medium team)
Large team / multi-team
Enterprise
GPU is only required if you also run a local LLM on the same host. For an OpenAI-compatible LLM running elsewhere (your own vLLM cluster, an LLM gateway, or a cloud API), CodeAlive itself runs CPU-only. Real footprint depends on repo size, indexing cadence, and concurrent users. Happy to size with you on a deployment call.
The Git servers, identity providers, and observability stack you already run.
Roadmap; available on Enterprise via custom deployment.
Today:Email/password + API keys with org-scoped RBAC
Self-hosted customers land here once they go live and approve a quote. No invented personas in the meantime. Want to be the first reference? Talk to us.
For now, see the homepage testimonials: Hauke Feddersen (grasbyte GmbH), Zhaksylyk Ualiyev (Esqadra Technologies), Alexander Kolotov (Blockscout), Sergey Loginov, and Sergey Sarafinovich. Those are real, published with permission.
You're not figuring out the Helm chart alone. We deploy with you.
Standard
Priority
Dedicated
Inside your VPC, against your LLM, talking to your internal Git server.