India's AI engineer. Your code never leaves the country.
A sovereign coding agent, served from India — or air-gapped inside your perimeter. It edits, runs, verifies and finishes real tasks in the editor your developers already use, and an independent referee stamps every claim.
One coding request, measured end to end
Editor → shield → the measured router — the tier is chosen from graded runs, never a guess. The completion streams back, an independent referee checks it, and a signed Trust Receipt rides home with it.
Every keystroke your team sends to a foreign coding cloud is your source code leaving India. Taksha is the answer — the same agentic workflow, on a model you own, on soil you choose.
Sovereign by design
Built in India, served from India. Your source code never crosses a border — data residency as a network fact, not a policy promise.
A model we own
Not a rented frontier API. A model family we own, fine-tune and operate — the intelligence layer answers to you, not a foreign cloud.
The independent referee
Every completion can be verified by a check the model never sees — your build, your tests — and stamped ✅ verified or ⚠️ unverified. No other coding assistant does this.
Measured, not marketed
We publish our benchmark table — including the cells we lose. Same tasks, same graders, against GPT-5.5, Claude and Gemini. Ask any vendor for theirs.
Works in your editor
One-click VS Code extension, plus any OpenAI-compatible client (Zed, Aider, SDKs, curl).
Governed by default
Per-developer keys and metering, org limits, SSO, audit — the controls BFSI and government actually require, built in.
Agentic coding — in your editor, on your model
Real multi-file, tool-using agent runs — nothing leaves your region.
The Taksha family — Fast · Pro · Max, and the Adhunik, Pariksha and Kriti specialists
Taksha is the engineering member of the AgentAnywhere sovereign model family — alongside Manthan · Kuber · Seva · Tatva · Astra · Sanjaya.
taksha-auto — routed by measurement, with the receipt in every response
Set your model to taksha-auto and each request is routed to the model that measures best on its task class — by the table below, generated from execution-graded battery ledgers, never hand-edited. Every response says what routed where, and why:
This is the live table the router reads — every cell is pass/n from graded runs, citing the ledger files behind it (hover a cell). A cell routes only with n ≥ 8; unclassified requests take the default tier, stated honestly, never a guess. A "—" means that tier hasn't accumulated n ≥ 8 graded runs on that class yet — Max is the newest tier, so its cells fill in as its ledgers land; we'd rather show you a dash than a number we didn't measure. Prefer thrift or the best measured rate? taksha-auto-cost and taksha-auto-quality tune the trade-off.
Served from India. In your VPC. An owned model layer.
For regulated buyers, that's the difference between a policy promise and a network fact — every completion served from a model we own, on hardware in your jurisdiction. Taksha is part of AgentAnywhere Sovereign — agentic AI that never leaves your borders, signed proof on every call.
Invite-only preview. Then your team. Then your VPC.
Straight answers
Does my source code leave India?
No. Taksha is served from India — or air-gapped inside your perimeter. Your source code never crosses a border: data residency as a network fact, not a policy promise.
How is Taksha different from other coding assistants?
Every completion can be verified by an independent check the model never sees — your build, your tests — and stamped verified or unverified. And routing is measured, not marketed: the routing table is generated from execution-graded battery ledgers and published live, including the cells we lose.
Which editors and tools does Taksha work with?
A one-click VS Code extension, plus any OpenAI-compatible client — Zed, Aider, SDKs, or curl — against one /v1 endpoint.
What is taksha-auto?
The measured router. Set your model to taksha-auto and each request is routed to the tier that measures best on its task class — a cell only routes with n ≥ 8 graded runs, and every response says what routed where, with the receipt.