Source-available · Business Source License 1.1 · self-hostable
VegaDūta: the complete suite for AI agents
Run governed agents on your laptop, your team’s server or Kubernetes — then reach them from your IDE, browser, phone and your own code. One repository ships the platform, the governance SDKs, the plugins, the apps and the installers.
On this machine, in minutes
curl -fsSL https://raw.githubusercontent.com/CD-LC/vegaduta/main/install.sh | sh
./scripts/vegaduta up --liteThe release ships pinned, signed container images - vegaduta up pulls them; no source build. Needs Docker with Compose v2 and an LLM provider key or a local Ollama — or none at all: the personal install ships a local model.
One suite, three ways in
VegaDūta is the platform you run, the SDKs you build on and the plugins and apps that carry agents into the tools you already use.
The platform
Agents with tools, an Assistant per user, workflows with human approval, retrieval and graph memory, omnichannel reach, guardrails, budgets and a sealed audit log. Try it in the browser, then put it on your own box.
The SDKs
Governance inside your own application: @Governed risk labels, human approval, budgets, idempotent effects and audit around any tool call. Java Agent SDK, Python Agent SDK, and generated TypeScript/Python REST clients.
Plugins and apps
IDE extensions for VS Code, JetBrains and Eclipse; Chrome and Edge companions; the Android app (beta); a Wear OS preview; a desktop agent; and VegaDūta itself as an MCP server your own client can drive.
From a laptop to a fleet
Same images everywhere: pulled from GHCR, pinned to the release, cosign-signed and attested — no build step unless you want one.
Personal
One person on a laptop: the Setup app, a bundled local model, sign-in with a passkey or PIN. Nothing to deploy.
See /downloadsLite
A small team's box: five containers, Keycloak sign-in, your own provider keys or host Ollama.
./scripts/vegaduta up --liteFull
The whole platform on compose: admin console, object storage, PgBouncer, every opt-in profile.
./scripts/vegaduta up --fullKubernetes
Umbrella Helm chart with scalable API, jobs, optional worker and every optional service as its own chart.
helm install vegaduta oci://ghcr.io/cd-lc/charts/vegadutaInside the platform
Every item below is a module in the repository, not a roadmap slide.
Agents and tools
Build agents on any supported LLM provider. Import OpenAPI 3.x / Swagger 2.0 specs as tools, register MCP servers per tenant, and attach Agent Skills (SKILL.md).
An Assistant for every user
Everyone gets their own Assistant agent with routines, threads, scheduled tasks and learned skills - on every install shape, personal to full.
Workflows with approvals
A node-based workflow engine with human-in-the-loop approval steps, hosted forms, and an audit log of who decided what.
Retrieval and graph memory
Ground agents in your own data with pgvector retrieval and graph memory on Apache AGE.
Omnichannel
Reach agents over web chat, voice, email, WhatsApp and other channels through adapters that share one approval model.
Build Studio
Draft an agent from a description, preview it, and publish it to a tenant - the same pipeline the hosted builder runs.
Guardrails and governance
Content-filtering guardrails, budgets, tenant isolation, and role-based access, enforced in-process on the chat hot path.
Artifacts, tables, schedules
Agents produce files and datasets, keep structured data tables, and run work on a schedule - all tenant-scoped.
Browser automation
An opt-in browser runner lets agents navigate, read and fill real pages under the same approval and audit rules.
Learning and Academy
Agents learn reusable skills with proposals you approve; the hosted Academy at academy.vegaduta.ai teaches the platform end to end.
Get VegaDūta
Pick your platform - every download resolves to the real artifact in the latest release.
macOS
VegaDuta-Setup-<version>.dmg
Universal app - checks Docker, pulls the signed release, opens the console. Gatekeeper warns once (unsigned build).
DownloadWindows
vegaduta-setup-windows-amd64.zip
Unzip and run the Setup exe; SmartScreen warns once. Installs the personal shape by default.
DownloadLinux
vegaduta-setup-linux-{amd64,arm64}.tar.gz
Both architectures. Untar, run the setup binary, done - or use the one-line terminal installer below.
DownloadTerminal
curl -fsSL https://raw.githubusercontent.com/CD-LC/vegaduta/main/install.sh | sh
./scripts/vegaduta up --liteKubernetes
helm install vegaduta oci://ghcr.io/cd-lc/charts/vegadutaEvery release artifact is listed in SHA256SUMS.txt - verify before you run. Images are cosign-signed and carry SBOMs. All releases.
Use it as a dependency
Run the platform, or embed the governance layer in your own application. They are separate things; pick by what your code is.
The Java Agent SDK wraps the tools you already have (Spring AI today, any framework through its framework-free core) so that a risky call waits for a human, a run has a budget, a retried turn never repeats a payment, and every action leaves an audit event tagged with the OWASP agentic risk it relates to.
<dependency>
<groupId>io.vegaduta</groupId>
<artifactId>vegaduta-agent-spring-boot-starter</artifactId>
<version>0.1.0</version>
</dependency>@Tool(description = "Refund an order")
@Governed(risk = RiskLabel.HIGH, requiresApproval = true, idempotent = true)
String refund(String orderId, long cents) { ... }Build it from a checkout or a release tag with git checkout v0.1.4 && mvn -f sdk/java/pom.xml install; Maven Central publishing is on the roadmap.
Java Agent SDK
io.vegaduta:vegaduta-agent-*
Governance inside your own Java application: tool risk classification, human approval, budgets, idempotent effects, guardrails, hooks, audit and telemetry around Spring AI tools. The core has no framework dependency.
Apache-2.0. Builds reproducibly from the v0.1.4 tag; Maven Central publishing is on the roadmap.
Python Agent SDK
vegaduta-agent (PyPI name reserved)
The same governance contracts for Python: @governed, ApprovalGate, budgets, guardrails, tamper-evident evidence chain - dependency-free, transports are injected ports.
Apache-2.0. In sdk/python-agent. First PyPI publish is on the roadmap.
TypeScript client
vegaduta-sdk
A dependency-free client for the platform API, generated from its OpenAPI document: agents, chat streaming, workflows.
In sdk/typescript. The package name is a placeholder until a registry name is reserved.
Python client
vegaduta_sdk
The same client for Python 3.9+, standard library only.
In sdk/python. Not published to PyPI yet.
Where it runs
Editors, browsers, phones and your own MCP client, each with where it really stands.
Code editors
- VS CodeAvailable
- JetBrains IDEsAvailable
- EclipsePreview
Browsers
- ChromeAvailable
- Microsoft EdgePreview
Phones and wearables
- Android appBeta
- Wear OS watchPreview
- iPhone appIn development
Your computer and other AI tools
- Desktop agentPreview
- MCP serverAvailable
Architecture and stack
A modular monolith core with optional services for async work and untrusted code.
- Core
- Java 21, Spring Boot, one Maven reactor of focused modules
- Consoles
- Next.js: tenant console and platform-admin console
- Data
- Postgres 16 with pgvector and Apache AGE; one schema per tenant
- Cache and queues
- Valkey; optional Temporal for durable workflows
- Identity
- Keycloak (OIDC), SSO, per-tenant isolation
- Services
- Python sidecar (ML), code-intel, sandboxed tool runner, MCP stdio runner
Open standards
Interoperable by default, so you are not locked to one client or vendor.
- MCP · client and server (Streamable HTTP, SSE, stdio)
- A2A · signed agent cards, JSON-RPC message/send and tasks
- AG-UI · server-sent-event agent runs
- OpenTelemetry · OTLP traces with gen_ai.* attributes
- OIDC · Keycloak or any compliant identity provider
- OpenAPI · 3.x and Swagger 2.0 import as tools
Built in the open
Licensed Business Source License 1.1 — free to run, fork and build on, including commercially; the restriction is offering it as a competing hosted service. Every version converts to Apache-2.0 four years after release. Issues, pull requests and discussion happen on GitHub.