Skip to content
Build, run and audit AI agents

Agents that operate for you, with autonomy you can prove

Describe the agent you need to Claude or Cursor, or write it in code. Knowledge, tools, WhatsApp or Discord, and every run traced with its cost.

$02,000 runs a month, free
1 APIask your assistant, or code it
Tracedevery run, with its cost
Ask or code

Ask for it.
Or code it.

Tell your assistant what the system should do and it builds it here. Or write the same calls yourself. Same API either way.

MCP server

Connect Claude, Cursor or VS Code and describe what you want built.

TypeScript SDK

@naturali/sdk — typed paths, bodies and responses.

CLI

@naturali/cli — one command per operation, scriptable in CI.

One contract

Every client is generated from the same OpenAPI, so none can do more than another.

your assistant, over MCPask

Create an agent that answers from our refund policy PDF. Ask me before it issues any refund over $100. Then put it on our Discord server.

@naturali/clishell
naturali create-agent-generation \
  --project-id "$PROJECT" \
  --agent-id "$AGENT" \
  --messages '[{"role":"user","content":"What is our refund window?"}]'
How it works

Configure.Connect.Prove.

EXAMPLEtrace_8f2c41 · 30 Sep 2026 · 09:41 UTC

Every run,
read back.

0 steps
  1. 01message received
  2. 02knowledge search · 3 chunks
  3. 03model turn
  4. 04tool_call lookup_order
  5. 05guardrail refund_limit · passed
  6. 06approval · a person said yes
  7. 07reply sent
One trace, read back
every step, in the order it ran
the config that served the run
Agent version
v0
split by the turns that spent it
Cost of the run
$0.000
Model turnsTool callsKnowledge retrievedGuardrail checksWhat it cost
After the fact

Questions you can answer later.

Why did the agent refund that order?

01

Traces

Every model turn and tool call in the order it ran

7–90days of traces kept, by plan
Answered by
Control

Autonomous,
not uncontrolled.

Guardrails gate what an agent may do, approvals put a human in front of the decisions that need one, and every request stays on the record.

Before it acts
4classes for every proposed tool call
A · executeB · execute if a guard passesC · ask a humanD · refuse
Pricing

Priced
on runs.

One run is one agent generation or one tool call, wherever it came from. Managed models draw on a prepaid balance; on your own credential, your vendor bills the tokens.

01

Free

Build one agent and prove it works on your data.

$0forever

no card needed

  • 2,000 runs a month
  • 3 projects · 1 channel
  • 1 GB indexed storage
  • 7-day trace retention
  • Evals included
  • Community support
Start free
02

Pro

Put an agent in production with a human in the loop.

$49/month

or $490 a year — two months free

  • 25,000 runs a month, then $2 per 1k
  • Unlimited projects · 10 channels
  • 30 GB indexed storage
  • 30-day trace retention
  • Guardrails, approvals and exceptions
  • Email support
Start with Pro
03

Business

Volume, a longer record and an audit log of every request.

$249/month

or $2,490 a year — two months free

  • 250,000 runs a month, then $1 per 1k
  • Unlimited projects · 100 channels
  • 100 GB indexed storage
  • 90-day trace retention
  • Audit log
  • Priority support
Start with Business
Members free on every planEvals on every planA downgrade waits for month end
Questions

Frequently
asked.

01What is naturali?
One API to build, run and prove AI agents. Give an agent knowledge and tools, put it on WhatsApp or Discord or call it from your product, and read back every generation it ran with what it cost.
02Where can an agent answer?
On WhatsApp and Discord through channels: you choose the agent, and the platform runs the conversation from the incoming message to the reply. Or call it from your own product over the REST API, the TypeScript SDK or the CLI.
03Which models can I use?
The managed catalog, with no vendor account, paid from a prepaid balance. Or your own credential, and your vendor bills the tokens. A model route tries its targets in order and falls through on a provider error, timeout or rate limit.
04Can Claude, Cursor or VS Code build it for me?
Yes. naturali is an MCP server: connect your assistant, approve it in your browser, and describe what you want built. Every operation in the API is a tool, so it reaches exactly what you would call yourself.
05How do I see what an agent did?
Every run is traced: model turns, tool calls, knowledge retrieved, guardrail checks and what it cost, with the agent version that served it.
06How much does it cost?
Free is $0 with 2,000 runs a month and no card. Pro is $49 and Business $249 a month; Enterprise is by contract. One run is one agent generation or one tool call.
Every billing question

Build the agent.
Keep the receipts.

Sign in, connect your assistant or open the quickstart, and run your first agent.

Free plan · 2,000 runs a month · no card