What INDRA is

Refineries, PSUs, defence-linked manufacturing and government offices produce a lot of routine but sensitive knowledge work — approval notes, board decks, engineering calculations, internal tooling code, review of scanned drawings and inspection reports. None of it can go through a cloud assistant, because the underlying material is confidential: P&IDs, financials, vendor negotiations, unreleased designs, internal correspondence.

So the work gets done by hand, or the confidential material quietly gets pasted into a public tool anyway.

INDRA is the third option: a self-hosted, air-gapped AI workbench that runs entirely on your own GPU server, and behaves like the assistants people actually want to use.

What it does

Runs fully on-premise. No external calls at any point. Not as a claim — the workbench carries a live sovereignty indicator on every screen, backed by a real egress log, and ships an interactive probe box so a sceptic can type any URL and watch it get blocked and recorded.

Picks the right model for the job. Multiple open-weight models loaded at once, selected automatically by what the task actually needs — a coding request routed differently from a document summary. New models drop in without redesigning anything.

Acts like an agent, not a chatbot. Plans multi-step work, calls local tools (file read/write, sandboxed code execution, document search), observes results, and iterates toward a real deliverable instead of answering once and stopping. The plan is visible while it runs, including when it gets revised mid-task.

Handles more than text. Scanned PDFs, handwritten notes, engineering drawings, P&IDs and photographs, read through on-device OCR and vision models.

Grounds answers in your own documents. A local knowledge base over your manuals, SOPs and past correspondence — with box-level citations back to the source page, so every claim can be checked.

Produces real files. Approval notes, Word/Excel/PowerPoint, working code, calculations with the steps shown — not just chat replies.

One-command run

Builds the backend and launches the desktop app together:

just run-ui

That is the whole thing — backend binary and UI, one command, on Linux, macOS and Windows. First run compiles the Rust backend, so give it a few minutes; after that it is near-instant.

Prerequisites: Rust, Node 20+, pnpm, CMake and a C/C++ toolchain. DEPENDENCIES.md lists exactly what to install per platform — including the Windows specifics that are easy to get wrong. A CI job installs only what that file prescribes on all three OSes and builds from scratch, so it stays honest.

Don't want a host toolchain at all? Run the backend in a container instead:

cp .env.docker.example .env    # set a secret
docker compose up --build

See docs/DOCKER.md.

Using it

The left rail is the whole app: Work (the conversation), Models, Sources, Memory, Trace, Sovereignty.

In the composer:

• / — run an installed skill
• // — hand the task to a specialist agent
• Auto / Manual — whether the agent acts on its own or asks before each action

Workspace folders are granted explicitly: pick specific folders (each read-only or write-with-approval), or grant full access deliberately. The model only ever sees what you've granted.

⌘K / Ctrl+K opens the command palette, which reaches everything.

Built on

Rust backend, Electron + React desktop client, and the Agent Client Protocol between them. Local inference through llama.cpp, with Ollama and OpenAI-compatible local endpoints also supported. Tools and integrations attach over the Model Context Protocol.

Project docs

• DEPENDENCIES.md — per-platform install requirements
• docs/DOCKER.md — containerized backend deployment
• AGENTS.md — contributor and build command reference
• CONTRIBUTING.md — how to contribute

---

Built for SIH Problem Statement 26117 — Sovereign On-Premise Agentic AI Workbench using Open-Weight Multimodal LLMs for Confidential Industrial Work, Mangalore Refinery and Petrochemicals Limited (MRPL).

_{Derived from the goose project. See LICENSE and NOTICE.}