BharatCode is an open-source, Rust terminal AI coding agent for India. It runs locally on your machine, works with open-weight and India-hosted models, and keeps your source code sovereign. It is Apache-2.0-licensed and free.

Yes. BharatCode is free and open source under the Apache-2.0 license. You only pay for whatever model provider you choose — or nothing at all if you run a fully local model with Ollama or LM Studio.

How do I install BharatCode?

Install with Homebrew (brew install arbazkhan971/tap/bharatcode), npm (npm install -g bharatcode-cli), or the install script (curl -fsSL https://raw.githubusercontent.com/arbazkhan971/bharatcode/main/install.sh | sh). Windows, go install, and source builds are also supported.

Does my code stay private with BharatCode?

Yes. BharatCode is local-first and sovereign by default. An offline mode rejects any non-local provider and disables network tools, and a tamper-evident audit log lets you prove your source never left the machine — built for banks, enterprises, and DPDP-regulated teams.

India's open-source AI coding agent

Build with AI.
Keep India in control

I'm building BharatCode as a sovereign alternative to Cursor, Claude Code, and Copilot — built for Indian developers, students, freshers, enterprises, and institutions using open-source and Indian AI models.

$brew install arbazkhan971/tap/bharatcode

also:npm i -g bharatcode-clicurl -fsSL …/install.sh | sh

14K+

YouTube developer community

2 years

free-student compute mission

Opt-in

privacy-preserving data flywheel

Join the BharatCode community Partner on compute

bharatcode — ~/rust-api · offline

$ bharatcode --offline

🔒 offline mode — code will not leave this machine

◆ model qwen2.5-coder:32b · provider ollama (local)

› add idempotency keys to the checkout handler

⊙ navigate references CheckoutService — 4 call sites

⊙ edit crates/payments/src/checkout.rs +24 −3

⊙ bash cargo test -p payments

✓ tests green · checkout handler now idempotent

audit ✎ 6 events recorded · bharatcode audit export

Local-first, Rust-powered, open-source, and model-flexible — built to help India own the AI coding workflow layer while keeping private code private by default.

Why this matters

AI coding workflow is becoming foreign-controlled infrastructure.

India should not only consume AI coding tools from the U.S. and China. India should build its own AI developer infrastructure — from compute and models to the coding agents developers use every day.

The workflow layer is strategic

AI coding agents are becoming the daily interface through which developers write software, choose models, route code, and learn engineering habits.

India must own its layer

The U.S. already dominates tools like Cursor, Claude Code, and Copilot while China is building its own AI model and developer ecosystem. India needs its own open developer infrastructure.

Students and freshers are the wedge

Millions of Indian learners should get access to AI coding workflows without being priced out or locked entirely into foreign platforms.

Product + compute mission

Help keep AI coding free for Indian students.

I'm looking for compute partners to help keep BharatCode free for Indian students and freshers for two years. Sponsored compute can turn BharatCode into a national AI coding adoption layer and help India build the data flywheel needed for stronger coding models.

Partner on compute View the Rust CLI

A local-first, open-source coding agent

Local-first Rust terminal agent that can run on-device and work with open-source models.
Model-flexible architecture for Indian providers, open-weight models, private deployments, and enterprise workflows.
Trust primitives for institutions: offline operation, auditability, data control, and predictable rupee-aware costs.

Compute → adoption → models

01Compute

02Free BharatCode access

03Student + fresher adoption

04Opt-in learning data

05Post-training open-source models

06Stronger India-first coding models

The goal is a stronger Indian AI ecosystem powered by sponsored compute, mass adoption, and privacy-preserving consent.

Data and trust

Private code stays private by default.

BharatCode will not train on private user code by default. Any training data contribution will be opt-in, consent-based, and privacy-preserving, with clear controls for users, colleges, and organizations.

Opt-in learning signals may include

Prompt-to-code edit trajectories

Bug diagnosis and patch workflows

Compiler error explanations and fixes

Test failure debugging paths

Student learning and problem-solving workflows

Hinglish and Indian English developer queries

Agent tool-use traces

Code review and documentation workflows

Community and partners

Built with India's developer community, not behind it.

BharatCode already has a 14K+ YouTube developer community that can help educate, onboard, and mobilize Indian developers around sovereign AI coding workflows.

Join the BharatCode community Partner with BharatCode

GPU creditsCloud creditsHosted inference capacityModel-serving infrastructureOpen-source model supportUniversity and fresher adoption programsEnterprise and public-sector pilots

Why BharatCode

Built on three non-negotiables

A terminal coding agent designed for India — sovereign by default, open-weight by choice, and shipped as a native Rust agent.

Sovereign by default
Offline + audited
Your source never has to leave the machine. An --offline mode rejects every non-localhost provider and disables web tools, while an append-only, tamper-evident audit log records each action — exportable proof for banks, enterprises, and DPDP-regulated teams.
Open-weight first
Your model, your choice
Open weights are first-class, not an afterthought — Kimi K2, DeepSeek V3 / R1, and Qwen Coder run locally or on Indian and Asian inference, at a fraction of frontier cost. Around 24 providers are wired in; closed models stay optional, never required.
Rust-native, single binary
Open source
A Rust-powered terminal coding agent built for predictable local workflows, open-source extensibility, and a rupee-aware cost ledger that keeps spend legible from the first run.

Capabilities

More than an agent — a sovereign toolkit

LSP-backed code intelligence and a 22-tool agent loop, wrapped in the offline mode and audit trail that keep your source where it belongs.

LSP-backed code intelligence
navigate · symbols · rename
Real Language Server power, not guesswork: go-to-definition, references, and declarations via navigate, plus symbols, rename, code actions, hover, format, and live diagnostics across your project.
Agent loop, 22 tools
bash · edit · patch · grep
An autonomous loop wired to a full toolset — read/write/edit/multiedit, multi-file patch, glob/grep/ls, bash, web fetch/search, todo, and background jobs — so it can drive a repo end to end.
Offline, air-gapped mode
--offline
Run with --offline and BharatCode rejects every non-localhost provider and disables web tools — a hard guarantee that your code will not leave this machine.
Tamper-evident audit log
bharatcode audit export
Every action is recorded to an append-only, tamper-evident audit log. Export the full trail with bharatcode audit export — compliance-grade proof of exactly what the agent did.
~24 model providers
sarvam · krutrim · ollama …
Indian-first by design: Sarvam and Krutrim sit beside Anthropic, OpenAI, Gemini (native), OpenRouter, DeepSeek, Kimi, Mistral, Groq, Cerebras, Together, Ollama, and LM Studio — around 24 in all. Bring your own, no lock-in.
Verified self-update
bharatcode update --apply
bharatcode update checks for releases; --apply downloads, verifies the checksum, and swaps the binary in place. Set auto_update: true to stay current automatically.
Benchmark eval harness
bharatcode eval
A built-in task suite measures real task success, steps-to-done, and recovery — so quality is something you can measure across every change, not just claim.
Shareable recipes
bharatcode recipes
Capture parameterized, repeatable workflows as recipes so your whole team runs the same tasks the same way every time — version-controlled alongside the repo.

Bring your own models

Your models. Your machine. Your country.

One agent speaks to 23 providers. Start fully local, pin to India-built models like Sarvam and Krutrim, and your source code never has to leave the country — or reach for open-weight gateways and frontier closed models when you choose to. Your data, your call.

LocalRuns fully on your machine — source code never leaves the device.
India / AsiaRegion-pinnable inference so data can stay in the country.
Open-weightOpen model families — often 10–20× cheaper than frontier closed models.

On your machine

Inference runs on the device. Source code never leaves the box — air-gap friendly by default.

Ollama
Local
Run open models fully on your own machine. Nothing leaves the box.
Open-weightdata stays on device
LM Studio
Local
Local model server with a friendly UI. Air-gap friendly.
Open-weightdata stays on device

Made in India

Sovereign, India-built inference on Indian infrastructure — keep traffic in-country.

Sarvam
India-built sarvam-m. Sovereign, made-in-India inference.
Open-weightIndia / Asia
Krutrim
Ola Krutrim — India-hosted models on Indian infrastructure.
Open-weightIndia / Asia

Open-weight gateways

Frontier-class open models — Kimi K2, DeepSeek, Qwen, GLM — often 10–20× cheaper than closed.

OpenRouter
One key, many models — route to open weights or in-region hosts.
Open-weightIndia / Asia
DeepSeek
DeepSeek V3 / R1 — strong open-weight coding & reasoning.
Open-weight
Moonshot / Kimi
Kimi K2 open weights — long-context agentic coding.
Open-weight
z.ai / GLM
GLM-4.6 open weights — capable, low-cost agentic coding.
Open-weight
Mistral
Mistral Large and Codestral — open-weight European models.
Open-weight
Groq
Open-weight models served at very high token throughput.
Open-weight
Cerebras
Wafer-scale inference for open weights at extreme speed.
Open-weight
Together
Hosted open weights — Qwen Coder, Llama, DeepSeek and more.
Open-weight
Fireworks
Fast inference for open-weight coding models.
Open-weight
DeepInfra
Low-cost hosted open weights — Qwen Coder and friends.
Open-weight
Nebius
Hosted open-weight Llama and Qwen at scale.
Open-weight
Novita
Open-weight Qwen and Llama through one hosted gateway.
Open-weight

Frontier, by choice

Closed frontier models for when you opt in — the same agent, your decision to make.

xAI / Grok
Grok models over an OpenAI-compatible endpoint.
Closed
Cohere
Command models via Cohere’s compatibility API.
Closed
Perplexity
Sonar models for web-grounded answers.
Closed
Google Gemini
Native Gemini 2.5 Pro / Flash, plus an OpenAI-compatible shim.
Closed
OpenAI
GPT and o-series frontier models via the native API.
Closed
Codex (ChatGPT)
Use your ChatGPT subscription via Codex OAuth — no API key.
Closed
Anthropic
Claude frontier models via the native Messages API.
Closed

Data stays in India. Local runtimes keep code on the device; region-pinned inference keeps it in-country. Built for banks, enterprises, and DPDP-regulated teams that can't ship source to a foreign cloud.

How it compares

An honest look at the landscape

Every tool here is good at something. BharatCode is the one built to prove sovereignty — an enforced offline mode and a tamper-evident audit log, on top of being open-weight-first and cost-aware in rupees.

Feature comparison of BharatCode, OpenCode, goose, and Claude Code.
Capability	BharatCodeOpen source	OpenCodeOpen source	gooseBlock · OSS	Claude CodeAnthropic
Offline / air-gapped modeMoatA flag that rejects non-local providers and disables web tools.	Yes--offline	No	No	No
Tamper-evident audit logMoatAppend-only, exportable record of every action the agent takes.	Yesaudit export	No	No	No
Data residencyMoatCan source code stay in India — local or India/Asia-hosted inference?	YesLocal-first	PartialProvider-dependent	PartialLocal supported	NoCloud (US)
Open-weight-firstOpen models (Kimi K2, DeepSeek, Qwen) as first-class defaults.	YesBy design	PartialSupported, not default	PartialSupported, not default	NoClaude only
Multi-providerSwap models freely — no single-vendor lock-in.	Yes~24 providers	YesMulti-provider	YesMulti-provider	NoAnthropic only
Cost-aware (INR)Built-in cost ledger that speaks rupees.	YesINR ledger	No	No	No
LanguageWhat the agent itself is written in.	Rust	TypeScript	Rust	TypeScript / Node
LicenseSource code license.	Apache-2.0	MIT	Apache-2.0	Proprietary

BharatCodeOpen source

This is us

Offline / air-gapped mode: Yes--offline
Tamper-evident audit log: Yesaudit export
Data residency: YesLocal-first
Open-weight-first: YesBy design
Multi-provider: Yes~24 providers
Cost-aware (INR): YesINR ledger
Language: Rust
License: Apache-2.0

OpenCodeOpen source

Offline / air-gapped mode: No
Tamper-evident audit log: No
Data residency: PartialProvider-dependent
Open-weight-first: PartialSupported, not default
Multi-provider: YesMulti-provider
Cost-aware (INR): No
Language: TypeScript
License: MIT

gooseBlock · OSS

Offline / air-gapped mode: No
Tamper-evident audit log: No
Data residency: PartialLocal supported
Open-weight-first: PartialSupported, not default
Multi-provider: YesMulti-provider
Cost-aware (INR): No
Language: Rust
License: Apache-2.0

Claude CodeAnthropic

Offline / air-gapped mode: No
Tamper-evident audit log: No
Data residency: NoCloud (US)
Open-weight-first: NoClaude only
Multi-provider: NoAnthropic only
Cost-aware (INR): No
Language: TypeScript / Node
License: Proprietary

Comparison reflects each project’s typical configuration and public licensing. Capabilities of fast-moving tools change — verify against upstream docs before standardising.

Get started

Up and running in one command

Install the binary, then point it at a hosted open-weight provider or a fully local model. Same agent either way — your data stays in India.

Step 01: Install the binary

A single static binary — no daemon, no telemetry, no Rust toolchain required. Pick whichever fits your machine.

homebrew

$ brew install arbazkhan971/tap/bharatcode

npm

$ npm install -g bharatcode-cli

curl · macOS / linux

$ curl -fsSL https://raw.githubusercontent.com/arbazkhan971/bharatcode-cli/main/install.sh | sh

windows · powershell

PS> irm https://raw.githubusercontent.com/arbazkhan971/bharatcode-cli/main/install.ps1 | iex

Step 02: Run with a hosted open-weight provider
Export one key and go. Open-weight models like DeepSeek V3/R1 run 10–20× cheaper than frontier closed models.
hosted provider
```
$ export DEEPSEEK_API_KEY=...        # or MOONSHOT_API_KEY, GROQ_API_KEY, etc.$ bharatcode
```
Step 03: Run fully local
No API key, no network round-trip — your source code never leaves your machine. Point at a local Ollama or LM Studio model.
ollama
$ bharatcode --provider ollama --model qwen2.5-coder:32b
lm studio
$ bharatcode --provider lmstudio --model qwen2.5-coder-32b-instruct

A few real flags

bharatcode --continueresume your most recent session.
bharatcode run --jsonheadless NDJSON event stream for CI and automation.
/goalkick off bounded autonomous work from inside the TUI.

View on GitHub Read the docs

AI coding workflow is becoming foreign-controlled infrastructure.

The workflow layer is strategic

India must own its layer

Students and freshers are the wedge

Help keep AI coding free for Indian students.

A local-first, open-source coding agent

Private code stays private by default.

Opt-in learning signals may include

Built with India's developer community, not behind it.

Built on three non-negotiables

Sovereign by default

Open-weight first

Rust-native, single binary

More than an agent — a sovereign toolkit

LSP-backed code intelligence

Agent loop, 22 tools

Offline, air-gapped mode

Tamper-evident audit log

~24 model providers

Verified self-update

Benchmark eval harness

Shareable recipes

Your models. Your machine. Your country.

On your machine

Ollama

LM Studio

Made in India

Sarvam

Krutrim

Open-weight gateways

OpenRouter

DeepSeek

Moonshot / Kimi

z.ai / GLM

Mistral

Groq

Cerebras

Together

Fireworks

DeepInfra

Nebius

Novita

Frontier, by choice

xAI / Grok

Cohere

Perplexity

Google Gemini

OpenAI

Codex (ChatGPT)

Anthropic

An honest look at the landscape

Up and running in one command

Step 01: Install the binary

Step 02: Run with a hosted open-weight provider

Step 03: Run fully local