Private beta · invites going out in batches · built in surat, india

Tell it in अपनी भाषा.
It ships with you.

An AI coding agent that plans, edits, and tests in a loop. Describe the task in Hinglish or 15 Indian languages — then review every change as a diff. Nothing touches disk until you stamp it.

Join the waitlist Watch a session →

375 tests passing·15 Indian languages + Hinglish·5 LLM providers·local-first

~/yuvadev — agent sessionstaged

$ yuvadev "auth.ts mein OTP login jodo"

▸ hinglish detected · loop engaged

▸ read src/auth/auth.ts (118 lines)

▸ staged: auth.ts +38 −4 · auth.test.ts +22

▸ tests 12 passed

⏸ awaiting your seal — review the diff, nothing is written yet

# your machine. your call.

how it works

One session, start to finish.

Ask in your language

Hindi, Hinglish, Tamil, Gujarati — or plain English. The agent detects it and gets to work; code always ships in standard English.

The loop runs

The model drives its own tools — read, edit, search, run — and iterates until the task is done. Your MCP servers plug straight into its toolbox.

The gate holds

Every write is staged in a virtual file system and validated. You see the full diff before a single byte reaches disk.

You decide

Approve, reject, or redirect. The same flow in the IDE and the yuvadev CLI — streamed tool calls, colored diffs, approval prompts.

mission-control — task #214loop

you · "सभी API endpoints में input validation जोड़ो"

▸ detected: hindi · plan drafted (8 endpoints)

▸ read src/api/routes_auth.py

▸ search "@router.(get|post)" — 8 matches

▸ mcp__github__list_issues (your MCP server, in the loop)

▸ edit staged: routes_auth.py +47 −3

▸ run pytest tests/unit — sandboxed

✓ 375 passed in 14.2s

⏸ approval gate — 1 file staged, diff ready [⌘↵]

you · approve

✓ committed to disk · vfs clean · session saved

capabilities

What it actually does.

Everything below is shipped and in the changelog — nothing aspirational on this list.

[01]loop

A real agent loop

The model drives the session: it reads, edits, searches, and runs commands in a loop until the task is done — not a fixed planner-coder pipeline.

[02]vfs · staging gate

Approval gate on every write

No file touches disk uninspected. Changes are staged in a virtual file system, validated, and shown to you as a diff before you approve them.

[03]mcp.json

MCP client built in

Point it at your MCP servers with a Claude Code-compatible mcp.json (stdio and HTTP) and their tools join the agent’s toolbox.

[04]cli

yuvadev in your terminal

The same agent as a CLI: streamed tool calls, colored diffs, approval prompts, and an interactive mode for back-and-forth work.

[05]i18n

Works in your language

15 Indian languages plus Hinglish, detected automatically. Describe the task in Hindi; the code ships in standard English.

[06]ollama · byok

Local-first by default

Run free local models with Ollama, bring your own Anthropic / OpenAI / DeepSeek key, or use YuvaDev Cloud. Your code stays on your machine unless you choose otherwise.

languages

बोलो अपनी भाषा में.

The agent understands 15 Indian languages plus Hinglish and English — detected automatically from how you type, script or no script. The code it writes always ships in standard English.

Hindi, Bengali, Telugu, Marathi, Tamil, Urdu, Gujarati, Kannada, Odia, Malayalam, Punjabi, Assamese, Maithili, Sanskrit, Nepali — plus Hinglish and English

>"लॉगिन पेज में validation जोड़ो"detected: hindi

>"navbar ne responsive banavo"detected: gujarati

>"இந்த bug ஐ fix பண்ணு"detected: tamil

>"yeh function refactor kar do"detected: hinglish

roadmap

Built in public, honestly.

This list mirrors the app's changelog — an item only moves to [x] when it has actually shipped.

Shipped (9)

[x]
Model-driven agent loop
The model picks tools (read, write, edit, search, run) and iterates until the task is done.
[x]
Virtual file system + approval gates
Every AI write is staged, validated, and diff-reviewed before it reaches disk.
[x]
MCP client
Claude Code-compatible mcp.json — stdio and HTTP servers plug their tools into the loop.
[x]
yuvadev CLI
Terminal agent with streamed tool calls, colored diffs, approval prompts, and interactive mode.
[x]
15 Indian languages
Hindi, Gujarati, Marathi, Tamil, Bengali and more — plus Hinglish — detected automatically.
[x]
Multi-provider LLM routing
Anthropic, OpenAI, DeepSeek, Ollama local & cloud, and YuvaDev Cloud, with health-aware fallback.
[x]
Project memory files
YUVADEV.md, AGENTS.md, and CLAUDE.md from your repo are injected into the agent’s context.
[x]
Self-healing CI loop
Agent runs tests, detects failures, and iterates fixes with convergence and budget guards.
[x]
Google sign-in
OAuth login that bridges the website and the IDE.

In progress (3)

[~]
Beta invites in batches
Onboarding the waitlist to private-beta builds, batch by batch.
[~]
Mission Control for parallel agents
Run and supervise multiple agent sessions side by side.
[~]
In-editor diff review
Approve or reject each staged change without leaving the editor.

Planned (2)

[ ]
Voice input
Speak your tasks in any supported Indian language directly to the agent.
[ ]
Team workspaces
Shared agent sessions and context for engineering teams.

pricing

Planned pricing.

⏸ nothing is billed today. YuvaDev is free during the private beta. These are the tiers we plan to launch with — waitlist members get 30 days of Pro when billing exists.

Free

$0/month

Local models, full agent.

Ollama local models
Full agent loop + approval gates
All 15 Indian languages + Hinglish
Community support

Join waitlist

Pro

Team

$49/month

For engineering teams.

Everything in Pro
5 team seats
Shared session history
Dedicated support

waitlist

Join the private beta.

Invites go out in batches as we onboard the waitlist. Early users get priority access, a direct line for feedback, and 30 days of Pro when pricing launches.

a note from the founder

I'm Jay Gautam, a CS (AI & ML) student in Surat, building YuvaDev in public. This month the agent loop, the MCP client, and the yuvadev CLI shipped — and beta invites started going out. If you build with it, I want to hear what breaks.

GitHub LinkedIn founder@yuvadev.dev

loading…

Tell it in अपनी भाषा.It ships with you.