Skip to main content

Self-hosted AI applications and agents

The AI platform you run.
Every answer cited. Every action approved.

Build assistants, agents and workflows on the models you choose, inside your own network. Nuvora grounds every answer in your documents, holds every consequential action for a second person, and keeps a record your auditors can verify.

$python3 -m nuvora.server --demono GPU · no pip install · no hosted model
nuvora.example.com
Nuvora playground with an answer and the evidence passages behind it
Sources citedevery passage with a content digest
Approval requiredauthor is never approver
Audit chain verifiedhash-chained, exportable

Your data stays home.

Prompts, documents and answers stay on infrastructure you run, and an exact host allow-list decides where any request may go.

Any model, your terms.

vLLM, Ollama or any OpenAI-compatible endpoint. Switch models without rewriting the app, and see every token against a budget.

Nothing acts alone.

Anything that writes, sends or spends waits for a different person to approve the exact action.

Proof, not promises.

Runs, guardrail decisions and approvals land in a hash-chained audit log you can export and check offline.

Meet Nuvora

Everything you need to put AI in production.
In one platform you control.

Use the best model for each job, on your endpoints.

Connect the models you already run and pick per task. Discover models automatically, set prices, and let routers start small and escalate only when an answer looks weak.

  • vLLM, Ollama, any OpenAI-compatible endpoint, or AWS-hosted models
  • Fabric and Gryvia presets with model discovery
  • An exact host allow-list: nothing reaches a host that is not on it
Connect a model →
Nuvora settings with model endpoints and integrations

Agents

Take agents to production with full control.

Every step an agent takes is bounded, screened and on the record. The moment it wants to change something real, a person signs off on the exact action first.

  1. Plan
  2. Call a tool
  3. Policy check
  4. Approve
  5. Act
  6. Record
  • A closed tool registry and a fixed step limit. No shell, no open internet.
  • Code runs in a Keep sandbox with no network, only after a different person approves the exact code.
  • Durable jobs survive restarts. Work interrupted by a failed worker is flagged for review, not lost.
  • Every step timed in a run waterfall, exportable to your tracing stack.
Nuvora approvals queue showing a pending action with its fingerprint

Do more with Nuvora

From idea to production, fast.

Ask your documents

Answers grounded in your policies, contracts and runbooks, with the source passage a click away.

Assistants that act

Open the ticket, issue the refund, update the record, after the right person approves.

Turn documents into data

Extract typed fields from forms and scans. Low-confidence results wait for a reviewer.

Images and audio

Read scans with OCR, transcribe calls, reason over images in chat, and generate images.

Automate workflows

Chain retrieval, models, reviews and handoffs into flows that pause wherever a human should decide.

Explain incidents

Ask why traffic dropped. Agents read Netra network evidence and cite it in the answer.

Why Nuvora

Everything a hosted AI platform does for you.
On ground you own.

Typical hosted AI platformNuvora
Where it runsThe provider’s cloud regionsYour hardware, your cloud account, or fully air-gapped
Where prompts goTo the provider’s serviceOnly to hosts on an exact allow-list you control
Who approves an agent’s actionWhatever confirmation steps you configureA different person, on the exact action, every time
What you can proveLogs in the provider’s monitoring serviceA hash-chained audit log you export and verify offline
Where agent code runsA managed runtime in the provider’s cloudA sandbox with no network, after approval
How model spend worksMetered by the providerYour endpoints, with a ledger and budgets per workspace

Hosted platforms bring managed model catalogs and regional compliance programs. Nuvora brings control.

Get started

Running in minutes. Yours for good.

Try it offline with Python 3.11 only, point it at a model you run, then deploy for the team with Helm or k3s. Existing OpenAI clients work unchanged, and every request passes guardrails, routing, budgets and the audit chain.

Try it
git clone https://github.com/zyvorai/zyvor-nuvora.git && cd zyvor-nuvora
export NUVORA_ADMIN_PASSWORD='choose-your-own-strong-password'
python3 -m nuvora.server --demo # http://127.0.0.1:8789

# k3s on your own host, HTTPS with a generated admin password
./scripts/deploy-remote.sh HOST USER
client.py
import os
from openai import OpenAI

client = OpenAI(
base_url="https://nuvora.example.com/v1",
api_key=os.environ["NUVORA_TOKEN"], # a scoped service token
)

reply = client.chat.completions.create(
model="auto", # or a model id, or "router:<id>"
messages=[{"role": "user", "content": "Summarize our refund policy"}],
)
print(reply.choices[0].message.content)

Put AI to work.
Keep it under control.

See Nuvora on your models and your documents, with a second person signing off on every agent action.

0.3.0 is an evaluation release. Zyvor Production License v1.0: free for evaluation, development and other non-production use; production needs a commercial license. The capability matrix says exactly what is and isn’t built.