SOVEREIGN CODING PLATFORM
Frontier coding. Sovereign by design.
Our own family of coding models — agentic in your editor today, governed by one control plane, and region-pinned so your source never touches a third-party cloud. Start free in your editor on our hosted infrastructure, or run the entire stack inside your own VPC. Sovereignty as a network fact, not a policy promise.
Install Taksha for VS Code, sign in, and code in two minutes — meet the model family →
SOC 2 · ISO 27001 · Independently assessed for HIPAA & GDPR · Signed Trust Receipt on every completion
THE PROBLEM
Your developers want AI. Your clients' contracts say no.
Every IT services firm, BPO, and regulated team hits the same wall. The productivity case for AI coding is settled — but hosted assistants route source code through clouds you don't control, in regions you didn't choose. For a team under client NDAs, data-residency clauses, or RBI and DPDP obligations, that isn't a feature trade-off. It's a breach waiting for an audit.
So the work goes one of two ways. It happens in the shadows — developers pasting client code into personal accounts — or it doesn't happen at all. Neither is a position you can defend to a client's security team.
THE SOVEREIGN THESIS
Most “private AI” is a promise in a contract. Ours is a fact on the network: the model runs where your code already lives, the request never crosses a border you didn’t draw, and every completion comes back signed.
You don’t trust a policy document to keep source code in-jurisdiction. You trust the route it took. Sovereign Coding makes the route the product — and hands you the receipt to prove it.
AGENTIC — TODAY, NOT SOMEDAY
Watch it modernize a legacy estate, in your perimeter.
Multi-file, tool-using, test-running agentic coding runs on our sovereign models with a 98K context (160K on Max) and streaming — a VB.NET → Angular migration is one prompt, not a fortnight.
- Reads the legacy file and plans edits across the whole feature — not one line at a time.
- Creates and updates multiple files, calls tools, and runs the test suite to prove the change.
- Signs a Trust Receipt and decrements your credits — every token accounted for, in-region.
FIG.SC1A real agentic run on the Pro model — reads, plans, writes, tests, signs.
OUR OWN MODEL FAMILY
A ladder, and specialists. One key for all of them.
Not a reseller catalog — a family we train and serve ourselves. A capability ladder for general work — Fast, Pro, Max — and specialists named for what they do: Adhunik for modernization, Pariksha for testing, Kriti for interfaces, Yantra for teams building agentic AI. One endpoint, one key: pick the specialist, or stay on Pro and let it take the work.
Pro
Our flagship coding model, fine-tuned on the migration and legacy-modernization work enterprises actually pay for. It plans across files, holds a large estate in context, and ships changes that pass the tests.
- VB.NET → Angular
- Legacy .NET / COBOL
- Repo-wide refactors
- Context
- 32K
- Interface
- Streaming · tool-calling
- On Team & up
- Pooled credits
Pro
5×The everyday default
The general model most teams live in: it edits, runs, verifies and finishes. 45/45 on our graded agentic battery with zero phantom completions — parity with a leading frontier model on the same tasks.
Max
20×Frontier-class · 160K
45/45 on the same graded agentic battery Pro passes, with zero phantom completions — a dedicated multi-GPU node for the largest codebases and longest context. Provisioned on request in about 15 minutes, as dedicated capacity rather than a shared tier.
Adhunik
5×Legacy modernization
COBOL, VB.NET and mainframe estates under migration doctrine: assess first, preserve behaviour exactly, prove equivalence, end with a source-to-target mapping.
Pariksha
5×Testing that bites
Unit, integration, E2E (Playwright, Selenium, Appium) and performance scripts — and it breaks your code to prove the suite goes red before claiming a pass. 89% mutation-kill on our graded battery.
Kriti
5×Interfaces that work · available
The UI specialist: components that are wired, accessible and behave under interaction — graded against a real DOM, not eyeballed. +15 points over the base engine across 8 UI task classes, n=128 per arm.
Yantra
5×For teams building AI · alpha
The agentic-AI specialist: agent frameworks, tool protocols, retrieval pipelines, evals and guardrails — for teams whose product IS the agent. Alpha · Target GA Q4 2026.
Running in-VPC and want a model of your own? Bring it. We certify customer-nominated and India-origin sovereign models into the same governed contract — same gateway, same Trust Receipt.
THE CODING STACK — SURFACES TO ENGINE
One engine. One family. Six tiers. Every surface you already use.
Read it bottom-up: a sovereign engine we pre-train, operate and fine-tune; the Taksha family standing on it; six tiers behind one key; and on top, the SURFACES layer — our own VS Code extension with the independent referee built in, Zed (verified), Aider and the CLI, or any OpenAI-compatible editor or SDK.
- Taksha for VS Code— our extension, independent referee built inInstall, sign in, code — and an independent referee checks every claimed success after the model claims it.
- Zed— verifiedVerified working end-to-end against the Taksha endpoint.
- Aider / CLITerminal-first agentic coding over the same OpenAI-compatible API.
- Any OpenAI-compatible editor or SDKPoint your editor or SDK at one base URL and one key — same models, same governance.
RIGHT TOOL FOR THE JOB
Name the job. We'll name the tier.
Six tiers, one key — and every line below is a claim we have already published and measured, not an adjective.
Agentic multi-step work — edit, run, verify, finish
Pro45/45 on our graded agentic battery, with zero phantom completions.
Sharpest single-shot generation & full-file rewrites
ProThe fine-tuned flagship — for the work where the first draft has to be right.
Legacy modernization — COBOL, VB.NET → modern stacks
AdhunikFlags undecidable business rules with greppable MIGRATION-QUESTION markers — audit-grade migration notes.
Test suites that actually catch bugs
ParikshaMeasured at parity with a leading frontier model on mutation-kill rate — 89% of seeded defects caught, zero invalid suites.
Largest codebases, 160K context
Maxenterprise · on-demandA dedicated multi-GPU node with a 160K context window.
UI building
Kritiin development · waitlist openThe UI specialist of the Taksha family — training corpus in generation now.
Not sure? Start with Yantra — the agentic default edits, runs, verifies and finishes, and you can switch tiers mid-project on the same key.
WHAT SHIPS TODAY
A platform, not a point tool.
A sovereign model layer, a policy gateway, and a signed Trust Receipt on every completion — the same control plane already running regulated agents in production. Keep your editor; consume it hosted, or run the whole stack inside your VPC.
AGENTIC & MULTI-FILE
Reads context, plans across files, and edits a whole feature in one run — the multi-step workflow enterprise migration actually needs.
TOOL-CALLING
The models call tools, run your test suite, and read results back — closing the loop instead of guessing at a diff.
32K CONTEXT · STREAMING
Hold a large estate in context and stream tokens as they generate. Real IDE latency, not batch.
GOVERNED GATEWAY
Every completion routes through one OpenAI-compatible endpoint where policy, redaction, and per-key metering are enforced at the point of the call.
TRUST RECEIPT
Each completion is signed — model, region, policy decisions, redactions — in the open AgentBOM format. Cryptographic proof for a client's auditor, not a verbal assurance.
CREDITS METERING
Usage is metered in credits (tokens × model weight) with per-key visibility, per-developer caps, SSO, and a full audit trail.
A REAL SESSION, REPLAYED
The model claims success. The referee has to agree.
This is a replay of a real field-test session — a boundary bug in a discount function, fixed and test-verified by Taksha in the editor. The finale is the part we insist on: an independent referee that re-runs the proof after the model claims success, and says so.
- Reads the file and finds the off-by-boundary bug — a 10% discount applied only above 1000 when the spec says at 1000 qualifies.
- Ships the minimal diff, then runs `node --test` itself — three tests, three passes.
- The referee then verifies the claim independently — exit 0, 1.2 seconds — and renders the verdict in the editor.
FIG.SC2Replay of a real Taksha session
TAKSHA FOR VS CODE
Our own extension. Install, sign in — and the referee rides along.
Taksha for VS Code is our branded, one-click extension. The enterprise developer installs it and signs in; there is no endpoint to configure, no model id to paste, no key to leak into a settings file. And the independent referee is built in — a check the model never sees, re-run after every claimed success.
- Referee verdicts in the editor — green when independently verified, orange UNVERIFIED with the evidence attached when a claim doesn't hold.
- Model identity in the status bar — you always know exactly which Taksha tier is serving you.
- Auto-update for pilot users — fixes and new tiers arrive without reinstalling.
Two-minute setup — works in VS Code, Zed, Aider, and any OpenAI-compatible editor.
Install
Add Taksha for VS Code from the console — one click.
Sign in
Authenticate with your AgentAnywhere account. No base URL, no model id, nothing to paste.
Code
Agentic, multi-file coding on the sovereign model — in your editor, in-region, from the first keystroke.
✅ Independently verified
`node --test` passed (exit 0, 1.2s) — verdict rendered by the referee, not the model.
⚠ UNVERIFIED
Model claimed “all tests pass”; the referee ran the suite and saw exit 1 — 2 of 3 assertions failed · full log attached.
Prefer another editor?
The platform speaks the OpenAI-compatible protocol your tools already expect. Point any of these at your hosted or in-VPC endpoint and they run entirely against sovereign models, with no context leaving your jurisdiction.
- Zed — verified
- Aider / CLI
- Continue
- JetBrains
- Any OpenAI-compatible base URL
TWO WAYS TO RUN
Same governance, delivered your way.
Consume Sovereign Coding hosted on our sovereign infrastructure, or run the whole platform inside your own VPC. The control plane, the models, and the Trust Receipt are identical either way.
HOSTED · SELF-SERVE
Run it on our sovereign infrastructure.
Region-pinned inference on our hardware, self-serve from the live console. Redeem an invite, get a key, and your team is coding in minutes — metered in credits, scaling by seat.
- Region-pinned, start free
- The live console at app.agentanywhere.ai
- Credits · pay-as-you-go overage · USD or INR
IN-VPC · AIR-GAPPED
Run the entire stack inside your perimeter.
Models, gateway, and governance deployed on your hardware, under your keys — hosted-grade or fully air-gapped. Nothing crosses a border you didn't draw; the same Trust Receipt on every call.
- Your VPC, your keys, your hardware
- Flat per-seat, unlimited usage
- SSO · audit · SLA · bring your own model
TRY IT HOSTED — OR RUN IT IN YOUR VPC
Start in your editor today. Move to your VPC when you're ready.
Metered in credits (tokens × the model's weight). Invite-only preview today; Team and Enterprise Dedicated as we scale. Enterprise pricing is finalised per customer in ₹, GST-compliant.
Pilot
Invite
Free during the invite-only preview. Seats are limited to what one machine can serve.
- Full agentic workflow in VS Code
- Independent verification built in
- One transferable invite to pass on
Team
Waitlist
Pooled seats for startups and product teams, metered per developer.
- Per-developer keys and metering
- Org console with per-dev caps
- SSO · usage dashboards
Enterprise Dedicated
₹ on request
Your own reserved serve — in your region, your VPC, or air-gapped. Annual commit.
- Dedicated GPU capacity, no sharing
- In-VPC / air-gapped deployment
- Uptime monitoring · audit · SLA
Max add-on
Metered
A frontier-class node with 160K context, provisioned on demand.
- 160K context window
- Dedicated multi-GPU node
- Billed per GPU-hour
Use the one-click Taksha for VS Code extension, or point Zed, Aider, Continue, or JetBrains at one OpenAI-compatible base URL. Drop in your key and you’re coding in two minutes.
BUILT FOR
For the teams who answer to someone else's auditor.
IT Services & SI partners
Deploy AI coding across your delivery bench without breaching a single client's source-control clause.
BPOs & GCCs
Give offshore teams modern tooling while every keystroke stays inside the contracted geography.
Regulated in-house teams
Banks, insurers, and health systems shipping their own software under RBI, DPDP, and HIPAA obligations.
WHY NOT JUST RUN IT YOURSELF
Anyone can download a model. Almost no one can prove what it did.
Ollama gets you a model on a laptop. It doesn’t get you a certification record, a policy plane, redaction, or a signed receipt your client’s security team will accept. Sovereign Coding is the difference between “we run AI locally” and “here is cryptographic proof of exactly which model touched your code, where, and under what policy.” That proof is the product.
FAQ
Frequently asked questions.
- Does Sovereign Coding send my source code to the cloud?
- No. Sovereignty is a network fact here: models and inference run either on our region-pinned infrastructure or inside your own VPC, under your keys. Source code, prompts, completions, repository context, and model weights never cross to a third-party cloud. The only thing handed out is a signed Trust Receipt you give to an auditor.
- Is the coding agentic — does it edit multiple files and run tests?
- Yes, today. The models are agentic and tool-using: they read context, plan across a whole feature, create and update multiple files, call tools, and run your test suite to verify the change — with 32K context and streaming. A VB.NET-to-Angular migration is one prompt, not a fortnight.
- What models does Sovereign Coding include?
- Our own family of coding models. A capability ladder — Fast, Pro and Max — where Pro is the everyday default that completed 45/45 on our graded agentic battery, and Max brings frontier-class reasoning with a 160K context on a dedicated node. Then the specialists: Adhunik for legacy modernization under a migration doctrine, Pariksha for tests it proves will fail on broken code, and Kriti for interfaces that are wired rather than merely written. Yantra, for teams building agentic AI, is in alpha. Usage is metered in credits (tokens times the model's weight: Pro, Adhunik, Pariksha and Kriti 5x; Max 20x, reflecting its dedicated multi-GPU node). Running in-VPC, you can also bring your own or a nominated India-origin sovereign model into the same governed contract.
- Which editors and IDEs does Sovereign Coding work with?
- The flagship way in is our own one-click Taksha for VS Code extension — install it, sign in, and you're coding on the sovereign model with nothing to configure, with an independent referee built in that verifies every claimed success. Zed is verified end-to-end, Aider and the CLI work over the same API, and any other OpenAI-compatible editor or SDK can point at your hosted or in-VPC base URL and run entirely against models inside your perimeter.
- How does the pricing work?
- Access is invite-only during the preview and free for pilot developers. Team, with pooled per-developer metering and SSO, is on a waitlist. Enterprise Dedicated gives you a reserved serve in your region, your VPC or air-gapped, priced per customer in rupees on an annual commit, and the Max node is available as a metered add-on. We publish honest numbers rather than a comparison table: usage is metered in credits, enterprise pricing is quoted against your actual seat count and deployment.
- What is a Trust Receipt?
- Every completion is signed — model, region, policy decisions, redactions — in the open AgentBOM format. It is cryptographic evidence of what ran, rather than a verbal assurance, so you can hand a client's auditor proof on demand.
- How is this different from just running Ollama or a local model myself?
- Downloading a model gets you inference on a laptop. It does not give you a certification record, a policy plane, redaction, per-key metering, SSO, or a signed receipt your client's security team will accept. Sovereign Coding is the governance layer between the model and the editor: cryptographic proof of exactly which model touched your code, where, and under what policy. That proof is the product.
- Does Sovereign Coding help with RBI, DPDP, HIPAA, or GDPR obligations?
- It is built for teams under data-residency and client-confidentiality constraints. Because inference stays inside your perimeter and every call is governed and signed, it supports deployments that must keep source code in-region. AgentAnywhere is SOC 2 and ISO 27001 certified and independently assessed for HIPAA and GDPR, and is designed to support India's RBI FREE-AI framework. It is a control that supports your compliance program — not a substitute for it.
SOC 2 · ISO 27001 · Independently assessed for HIPAA & GDPR · Region-pinned & in-VPC · Designed to support India's RBI FREE-AI framework · Signed Trust Receipt (AgentBOM)
DESIGN-PARTNER PROGRAM
Hosted is live. In-VPC pilots are onboarding now.
Start free on the hosted platform today. For in-VPC and air-gapped deployments, we’re onboarding a small group of services firms and regulated teams to shape the integration against real estates. Design partners get priority model tuning, hands-on deployment, and pilot pricing.
What we’ll ask
- Work email
- Company
- Role
- Developers on the bench
- Primary IDE
- GPU footprint (optional)
We reply within two business days.
Give your developers AI. Give your clients proof.
Start free on the hosted platform in two minutes — or bring your bench and your toughest client clause, and we'll scope an in-VPC deployment against your stack.