Skip to content
SentientOne

AI for the product you already have

Add AI to your product this week

Plug a production-ready AI agent into your app, website or workflows with one API call. Grounded in your docs, connected to your tools, and every answer traceable. No ML team required.

Start freeBook a setup call

200 free credits when you sign up. No card needed.

Works with your stack

  • OpenAIOpenAIModels
  • AnthropicAnthropicModels
  • Google GeminiGoogle GeminiModels
  • GroqGroqModels
  • Model Context ProtocolModel Context ProtocolTools
  • Cohere RerankRetrieval
  • pgvectorpgvectorRetrieval
  • OpenTelemetryOpenTelemetryTracing
  • SlackSlackWorkflows

What you can build

Build these in no time

Pick one, follow its step-by-step setup, and it's live in your product the same week.

How it works

One API call in, a grounded answer out

Your product sends a message. SentientOne checks the key and credits, searches your docs, calls your tools and streams back an answer with its sources: the same six stages for every channel. Timings are medians from our latest benchmark.

How SentientOne answers a question. A user message arrives from the website widget, the REST API or a workflow Agent step, with its conversation history, the agent's settings and a request id. 1: the key, the per-agent rate limit (60 requests a minute) and the credit balance are checked and one credit is reserved. 2: the answer cache is checked; a repeated first question is answered in about 46 milliseconds. 3: on a miss, knowledge is searched with pgvector and Postgres full-text search, fused with reciprocal rank fusion and reranked by Cohere rerank-v3.5, about 450 milliseconds, and the top five chunks go on. 4: the model may call tools on your MCP servers, up to eight rounds, with the results going back into its context. 5: your agent's model (for example GPT-5.5, Claude Sonnet 4.6, Gemini 2.5 Flash or Llama 3.3 70B) writes the answer with a grounding rule. 6: the answer streams back as server-sent events. The first token arrives in about 600 milliseconds, or about 1.1 seconds with knowledge. The response cites its sources and leaves a trace with its spans, tokens, credit and request id.

What you get

Everything behind that API call, already built

Skip stitching together a model API, a vector database, a prompt tool, a chat widget and a tracing service. Configure an agent once and it answers in your app, on your website and in your workflows, with every answer traceable to its sources.

See the platform map
app.sentientone.ai/agents/support-agent

Support agent

Claude Sonnet 4.6 · Active

Live
SettingsPromptKnowledgeToolsDeployVersions

Playground

Can I return a jacket I've worn once?

2 sources · tool get_order

Worn items can't be returned, but they're covered by the 12-month warranty for faults 1

612 ms first token1,894 tokens1 credit

An agent's page: the draft system prompt beside the Playground, where a reply shows its sources, a tool call, time to first token, tokens and its one credit.
Learn more about build an agent
Learn more about ground it in your content
Learn more about automate the work around it
Learn more about ship it anywhere
Learn more about see every request

Fits your stack

Works with what you already have, and the model you choose

Bring your docs, your APIs (as MCP tools) and your own OpenAI, Anthropic, Gemini or Groq key. Your product calls one endpoint, embeds one script tag or fires a webhook.

What you bring

SentientOne

One account, every part of an agent

Tracing: every request, every span, tokens and timing

Where it answers

Production-ready

Built for real customers, not for demos

The parts that decide whether AI survives contact with your users: search that finds exact product codes, keys safe for the browser, limits, retries and a trace for every reply.

  • Hybrid RAG with rerank

    pgvector (HNSW) and Postgres full-text search on every question, fused with reciprocal rank fusion and reranked by Cohere rerank-v3.5.

  • MCP tool calling

    Agents call tools on your MCP servers mid-answer, up to 8 rounds, with every call streamed to you and traced.

  • Server-sent events

    Replies stream token by token, with sources, tool calls and token usage as typed JSON events.

  • Scoped keys

    A secret sk-so- key for servers, a publishable pk-so- key per agent for websites, locked to up to 50 allowed sites.

  • Workflows with guardrails

    15 block types, PII and prompt-injection checks, signed webhooks, draft and publish.

  • Versioned prompts

    Every change to an agent's prompt or settings is a version with a diff. Restore any of them in one click.

  • OpenTelemetry tracing

    GenAI semantic conventions, personal details redacted, kept 90 days, searchable by agent, model, status and error.

  • Answer cache

    Repeat first questions are answered from a 15-minute cache, cleared the moment knowledge changes.

Grounded in your data

Answers from your docs, not the model's memory

Every question searches your content by meaning and by exact wording, and a reranker picks the best passages. Answers cite them, and SentientOne never trains a model on your content.

  • Chunks of 400 words with 40 words of overlap, embedded with OpenAI text-embedding-3-small (1536 dimensions).
  • Vector (pgvector, HNSW) and keyword (Postgres full-text, GIN) results fused with reciprocal rank fusion, 24 candidates reranked by Cohere rerank-v3.5.
  • The top 5 passages go to the model; a sources event tells your app which ones, with titles, URLs and scores.
  • Agents with knowledge get a grounding rule in their instructions, so they answer from your content.
How knowledge works
PDFUp to 50 MB
WebsiteUp to 200 pages
SnippetA fact or rule
Q&AExact answers
Chunk and embed400-word chunks, 40 overlap, OpenAI text-embedding-3-small (1536 dims)
Vector searchpgvector, HNSW index
Keyword searchPostgres full-text
FuseReciprocal rank fusion (k = 60), 24 candidates
RerankCohere rerank-v3.5
AnswerTop 5 passages to the model, sources streamed to you

For your engineers

One endpoint, every event typed

Your engineers integrate with one POST and a stream of typed JSON events: sources, tool calls, tokens and errors, each with a type they can rely on. No SDK to adopt, no model code to maintain.

Streaming events reference
$ curl -N https://api.sentientone.ai/v1/chat/stream \
  -H "Content-Type: application/json" \
  -H "X-Api-Key: $SENTIENTONE_API_KEY" \
  -H "X-Agent-Id: $AGENT_ID" \
  -H "X-Request-Id: req-2846c8227a41" \
  -d '{"message": "Can I get a refund on my annual plan?"}'

type StreamEvent =

11 events

  • received| { type: "meta"conversation_id: stringtrace_id: string}
  • received| { type: "sources"sources: { index: number; id: string; title: string | null url: string | null; filename: string | null source_type: string; score: number; snippet: string }[]}
  • received| { type: "tool_call"id: string; name: stringinput: Record<string, unknown>}
  • received| { type: "tool_result"id: string; name: stringoutput: string}
  • ×6| { type: "delta"content: string}
  • received| { type: "done"conversation_id: string; trace_id: stringusage: { input_tokens: number; output_tokens: number total_tokens: number; cost_usd: number }cached?: true; timing?: Record<string, number>}
  • | { type: "error"code: string // e.g. TIMEOUT, LLM_RATE_LIMITEDmessage: string; request_id: string}
Time to first token
612 ms
Output tokens
30
Credits
1
Request id
req-2846c8227a41

A request to POST /v1/chat/stream with X-Api-Key, X-Agent-Id and X-Request-Id answers with server-sent events, each a data line holding JSON with a type: meta (conversation_id, trace_id), sources (the passages used), tool_call and tool_result (an MCP tool and what it returned), delta (each piece of the answer), and done (conversation_id, trace_id and token usage); a failed reply ends with error (code, message, request_id) instead.

Get started

From sign-up to live in your product in four steps

  1. 01

    Add your LLM key

    Your own OpenAI, Anthropic, Gemini or Groq key, encrypted with AES-256-GCM. You pay the provider directly.

  2. 02

    Point an agent at your docs

    Start from one of 15 templates. Upload PDFs (up to 50 MB), crawl your help centre, add exact Q&A.

  3. 03

    Test it with real questions

    The Playground shows each reply's sources, tool calls, latency and tokens before anyone else sees it.

  4. 04

    Ship it in your product

    One API call from your backend, one script tag on your site, or a step in a workflow.

Start freeFollow the quickstart

200 free credits when you sign up. No card needed.

Safe by design

Safe to put in front of your customers

Your keys are encrypted, browser keys only reach one agent from your domains, outbound calls can't reach private networks, and every reply is traced. No badges, just the mechanisms.

  • Keys and secrets

    • Provider keys, MCP credentials and webhook secrets encrypted with AES-256-GCM
    • Saved keys are never sent back to the browser
    • Secret keys refused on the widget path

  • Outbound calls

    • MCP, HTTP and webhook calls only to public addresses
    • Connections pinned to the checked address, no redirects
    • Workflow webhooks signed with HMAC-SHA256

  • Access

    • Two-step verification with an authenticator app
    • Breached passwords refused at sign-up
    • Database row-level security on; all access through the API
Security in detail

Predictable costs

Pay per reply, not per seat

200 free credits to start, no card. One credit per reply, packs from $10, credits never expire and every feature is included. Your model provider bills tokens on your own key.

Credit packs

Prices in US dollars. Credits never expire, every feature is included, and there are no plans or seats.

  • On sign-up

    200 credits

    Free, once per account. No card needed.

  • Pack

    $10 USD

    1,000 credits, 1¢ each

  • 10% less per credit

    $45 USD

    5,000 credits, 0.9¢ each

  • 20% less per credit

    $200 USD

    25,000 credits, 0.8¢ each

  • 30% less per credit

    $700 USD

    100,000 credits, 0.7¢ each

  • Enterprise

    Custom hosting on your own infrastructure

    Talk to us

Auto top-up (off until you turn it on) buys a pack of your choice when the balance falls below a threshold, with a monthly cap. Agents use your own OpenAI, Anthropic, Google Gemini or Groq key, so model usage is billed by the provider to you, at their prices.

What uses credits

What each action costs in credits
An agent reply (Playground, widget, API or a workflow Agent step)1 credit
A page of a PDF you add to knowledge5 credits
A page crawled from a website5 credits
A snippet or a Q&A pair1 credit
A Classify or Extract block, or a Guardrail check with an AI-judged rule, in a workflow1 credit
A live workflow run with an HTTP, Webhook, Slack or Email block1 credit per run

The run credit applies only to live runs with outbound blocks, not to test runs in the designer. A reply that fails is refunded; so is a page that fails to ingest. Deleting knowledge doesn't refund it. At zero, requests get HTTP 402 OUT_OF_CREDITS until you top up.

Estimate your first month

3,000

Every reply from any channel

100

PDF or website pages, 5 credits each, paid once when added

Credits needed
3,500
After the free 200
$40

Cheapest with 4 × 1,000 credit packs, 700 credits left over. Your LLM provider bills model usage separately.

Start with 200 free credits

Help centre

Everything your team needs to ship, written down

Guides, API reference and troubleshooting, taken from the product itself. Search it, or start with the quickstart.

FAQ

Questions product teams ask

Something else? The help centre goes deeper.

Do we need an ML team?
No. You configure the agent in the dashboard (prompt, model, knowledge, tools) and your engineers call one endpoint or paste one script tag. There's no model to train or host.
Is it trained on our data?
No. Your docs are indexed and searched for each question (retrieval), and the best passages go to the model with the question. SentientOne doesn't train models on your content.
What exactly is SentientOne?
An AI platform for building and running agents. You configure an agent (prompt, model, knowledge, tools) in the dashboard, then use it through a streaming REST API, a website widget or a workflow. Every request is traced, and you pay per reply with prepaid credits.
Do I need my own model provider key?
Yes. Each agent runs on your OpenAI, Anthropic, Google Gemini or Groq key, stored encrypted. The provider bills you for tokens at their prices; SentientOne charges credits for replies and knowledge. You can also pass a key for a single request in the X-Llm-Key header.
What does one credit buy?
One agent reply, from any channel. Adding knowledge costs 5 credits per PDF or website page and 1 per snippet or Q&A pair. New accounts get 200 free credits, no card, and packs start at 1,000 credits for US$10. Credits never expire.
Is it really hybrid search, or just a vector lookup?
Both. Each question runs a pgvector search and a Postgres full-text search, the two lists are merged with reciprocal rank fusion, and Cohere rerank-v3.5 reranks the candidates. The top 5 passages go to the model, and the stream tells your app which sources it used.
Can I put an agent on my website without exposing my API key?
Yes. The widget uses a publishable pk-so- key, never your secret key (secret keys are refused on the widget path). It only chats with its own agent, only from the websites you allow, within 120 requests a minute per key and 12 per visitor.
How do I debug a bad answer?
Open its trace. It shows what was retrieved and with which scores, each model round, every tool call with its input and output, tokens and the error code if there was one. Send your own X-Request-Id and you can find the trace from your logs.
What does SentientOne store, and how is it protected?
Your agents, knowledge (text chunks and their embeddings), conversations and account. Provider keys, MCP credentials and webhook secrets are encrypted with AES-256-GCM before they're stored. Traces are kept 90 days with emails, keys and phone numbers redacted.
Is there a self-hosted version?
Not as a download. Enterprise customers can talk to us about custom hosting on their own infrastructure.

Add AI to your product this week

Start with 200 free credits, no card. Add your LLM key, point an agent at your docs, and call it from your app, your website or a workflow. Want help? Book a setup call.

Start freeFollow the quickstart