VedaCODE
Sovereign AI Software Engineering Platform
Self-hosted AI software engineering with maker-checker verification: requirements become planned, approved, sandbox-tested code changes on private NVIDIA GPUs, with no source code leaving your network.
- Runs open-weight coding models on your own GPUs — no code uploads
- Independent checker model reviews every change and runs the tests
- Contract-bound plans: nothing changes before four-eyes approval
- Developer CLI and VS Code extension, governed from VedaCODE Studio
Overview
Regulated engineering teams want agentic coding speed but cannot stream source code to external services, and single-pass code suggestions arrive untested. VedaCODE is a self-hosted AI software engineering platform that turns requirements into planned, approved, executed and verified code changes on private NVIDIA GPUs.
A maker model writes code and an independent checker model reviews it; builds run inside Bubblewrap sandboxes against unit tests and linters; and every build is bound to a hashed, human-approved plan. Developers work from a standard-library CLI or VS Code, while administrators govern models, routing, quotas and audit trails in VedaCODE Studio.
Key capabilities
-
Self-hosted models
Runs open-weight coding models locally through Ollama with no outbound telemetry or external code uploads.
-
Maker-checker review
An independent checker model evaluates the maker's code, checks syntax and runs tests; both models stay resident in GPU memory to avoid reload latency.
-
Contract-bound plan approvals
Each plan defines acceptance criteria, assumptions and open questions; execution is bound to the SHA-256 hash of the approved plan, with four-eyes sign-off.
-
Sandboxed builds and tests
File edits and test commands run inside an isolated Bubblewrap sandbox; unapproved material changes are escalated to human review.
-
Automated repair
N-round automated patching when tests or linters fail.
-
Failure integrity
A build that does not pass verification ends as FAILED with full step logs; broken code is not reported as complete.
-
Complexity routing
Prompts are scored for complexity (levels 1–5): routine tasks go to fast models, complex tasks to deeper reasoning models, with fallbacks and SLA response budgets.
-
Developer CLI
Standard-library Python CLI with /fix, /review, /test and /pr commands, working inside existing repositories.
-
VS Code extension
Served by the on-premises server; question, review and trigger verified builds from the IDE.
-
VedaCODE Studio
Web console governing builds, models, routing, tenants and audit trails.
-
FinOps and GPU fleet
Token and GPU-time metering per model and organisation, spend caps, fair-share queues and enrolled remote GPU workers.
-
CI/CD and issue automation
Failing CI test runs and issues trigger automated sandboxed builds.
Architecture
VedaCODE is installed on your own server by a single setup script. Developers reach a FastAPI control plane over HTTPS; builds run in Bubblewrap sandboxes and models are served by Ollama on local NVIDIA GPUs, with optional remote GPU workers.
-
Client access (HTTPS)
- Developer CLI (vedacode)
- VS Code IDE extension
- VedaCODE Studio web console
-
Control plane
- API engine
- Build lifecycle
- Complexity router
- WAF and guardrails
-
Data and runtime
- PostgreSQL 16 + pgvector code search
- Redis 7 task queues and cache
- Ollama runtime
- Bubblewrap sandboxed builds
-
Compute
- Local NVIDIA GPUs (24 GB+ VRAM recommended)
- Optional remote GPU workers (vedacode-worker)
Technical specifications
| Version | VedaCODE 0.1.0 |
|---|
| Control plane | FastAPI API, Studio web console, build engine, routing, tenancy, WAF and TLS |
|---|---|
| Data stores | PostgreSQL 16 with pgvector; Redis 7 (bound to 127.0.0.1) |
| Runtimes | Ollama (16k context, 2 parallel, flash attention); vLLM and llama.cpp adapters |
|---|---|
| Logical models | VedaCode-Fast, VedaCode-Coder, VedaCode-Deep plus custom, each with fallbacks |
| Example maker / checker pair | gpt-oss:20b + qwen2.5-coder:7b resident together on a 24 GB GPU |
| Policy graph | Conditions, logical gateways, SLA response budgets with failover, load balancing, circuit breakers, maker-checker review, Test Lab (complexity 1–5) |
|---|
| CLI | Standard-library Python 3.9+ on Linux, macOS and Windows; /fix, /review, /test, /explain, /pr, code search, skills |
|---|---|
| IDE | VS Code extension served by your server |
| Transport | TLS 1.2+ with ECDHE-only ciphers; self-signed P-384 certificate with auto-renewal or your own; HSTS with CA-signed certificates |
|---|---|
| Identity | Six tenant roles, members-only projects, OpenID Connect SSO (PKCE), AES-GCM sealed secrets |
| Sandbox | Build commands run in a sandbox (bubblewrap when available) inside the build workspace |
|---|
Deployment
-
Self-hosted / on-premises
Runs on your own servers and GPUs; source code stays inside your network.
-
Air-gapped Source-reported
Operates without outbound connectivity.
Requirements
| Operating system | Ubuntu 22.04, 24.04 or 26.04 (x86_64) |
|---|---|
| GPU | NVIDIA GPU, 24 GB or more VRAM recommended |
| RAM / disk | 16 GB RAM, 100 GB disk |
| Network | Internet access during setup |
| Installer | setup.sh: NVIDIA driver, Docker, NVIDIA container toolkit, Ollama, PostgreSQL 16 + pgvector, Redis 7, systemd units, models sized to the GPU |
| Developer machines | Python 3.9+ (Linux, macOS, Windows); VS Code for the extension |
| Evaluation | 30-day air-gapped proof of concept with pilot developer squads Source-reported |
Supported integrations
-
GitHub, GitLab
Source control and CI
Issues and failing CI runs become builds; verified builds become pull / merge requests.
-
Jira
Issue tracking
Issues can trigger builds.
-
Slack, Microsoft Teams, webhooks, e-mail
Notifications
Build and approval notifications.
-
Visual Studio Code
IDE
Extension served by your server.
-
OpenID Connect
Identity
Single sign-on.
Security & governance
-
Built-in WAF
Protects endpoints from scanners, path traversal and secret probing, with Standard, Strict and Under Attack profiles and auto-escalating rate limits.
-
Secret-stripping guardrails
Redacts API keys, credentials and PII before they reach prompts or logs; rejects builds that introduce hard-coded secrets.
-
Multi-tenancy and RBAC
Six tenant roles and members-only projects; every query and repository action is scoped to the caller's organisation.
-
Sealed credentials
AES-GCM sealed credentials for integrations; the CLI never uploads local credential files.
-
Audit logging
L3/L7 request audit logging and full build step logs.
Use cases
-
Regulated enterprises
Verified issue resolution
Turn issues and failing CI runs into reviewed, tested pull requests without code leaving the network.
-
Engineering teams
Multi-model code review
Independent checker review of every change with /review.
-
Platform and security teams
Governed AI adoption
Central control of models, quotas, routing and audit trails across teams.
Documentation & developer resources
CLI Usage
Security Controls
Architecture
Open the VedaCODE documentation
Downloads
- VedaCODE — Technical & Customer Deck
VedaCODE technical and customer briefing (PowerPoint, 10 slides).
Frequently asked questions
Does any source code leave our network?
No. Models run on your own GPUs through Ollama, and the CLI and VS Code extension talk only to your VedaCODE server.
What stops an AI agent from changing code without approval?
Every build is bound to the SHA-256 hash of a human-approved plan. Changes run in a Bubblewrap sandbox, and material changes outside the plan are escalated for review.
What hardware do we need?
An NVIDIA GPU with 24 GB or more VRAM is recommended, so the maker and checker models can stay resident together. Additional GPU servers can be enrolled as remote workers.
Items marked Source-reported are as stated in VedaCLOUD.AI product material and have not been independently verified.