Production agents
in days, not months

Define what your agent should do. SPIRITT runs it on managed infrastructure and any model, with sandboxing, long-running sessions, scoped permissions and tracing built in

support-agentManaged agent
Live
Task

Resolve order and refund tickets end to end, in the customer's language

Tools
ZendeskShopifyStripeGmail+ Your API
Guardrails

Refunds over $200 wait for a human

Never share payment details

Escalate legal and safety questions

Success criteria

Ticket resolved, customer confirmed, within policy

Model mixtureDeploy
ModelsAny provider, switch anytime
Mixture on
SPIRITTClaudeGPTGeminiGrokKimiDeepSeekMuseQwenGLM
Each step goes to the model that handles it best
Read the ticketGemini
Plan the resolutionClaude
Write the replyGPT
Check against policyDeepSeek
New model out? Swap it in, the agent stays the same
Program searchTuning against your success criteria
Searching

Default harnessBaseline

Tools reshapedBetter

Prompts tunedBetter

Custom harnessSelected

What changed

+Added tool order_history

~Rewrote the refund policy prompt

~Context keeps the last three orders

−Removed generic web search

Fitted to support-agentPromote
Run · ticket 4821Every tool call and decision
Running
read_ticket

lookup_order

check_policy

draft_reply

issue_refund

DecisionA $240 refund is over the $200 limit, so the agent asked Ana before calling Stripe
Waiting for approvalApprove

Checkpoint saved, the run survives disconnects

Every step inspectable, end to endSession 2h 14m
PermissionsScoped to this agent
Enforced

support-agentOwn identity · managed credentials

zendesk.read_ticketAllow
shopify.lookup_orderAllow
gmail.send_replyAllow
stripe.issue_refundAsk
crm.delete_contactDeny
Audit log
10:42 issue_refund · asked Ana · approved
10:41 draft_reply · allowed
10:40 lookup_order · allowed
research-leadDirecting 4 agents
Iterating
Competitor pricingDone
Review miningRunning
Market sizingRunning
Draft the briefWaiting
Outcome · a brief sales can use today

Every claim has a source

Covers the top five competitors

Fits on two pages, rewriting

Iterates until every criterion passes
SPIRITTClaudeGPTGeminiGrokKimiDeepSeekMuseQwenGLMSPIRITTClaudeGPTGeminiGrokKimiDeepSeekMuseQwenGLM

Everything between prototype
and production, handled

Program searchYour criteria

Default harnessBaseline

Tools reshapedBetter

Prompts tunedBetter

Custom harnessSelected

Custom harness

Program search tunes the harness, tools and prompts to your workload

RequestResponse
{
  "agent": "support-agent",
  "input": "Where is order 4821?"
}
{
  "status": "resolved",
  "reply": "It ships today, tracking…"
}

One clean API

Every agent ships behind an endpoint your product can call

Trace · ticket 4821Every call
read_ticket

lookup_order

check_policy

draft_reply

issue_refund

decision · over the $200 limit · asked a human

Execution tracing

Inspect every tool call, decision and failure, end to end

Permissions · support-agentScoped
zendesk.read_ticketAllow
shopify.lookup_orderAllow
stripe.issue_refundAsk
crm.delete_contactDeny
10:42 issue_refund · approved by Ana

Scoped permissions

Allow, ask or deny every tool, so real access never means lost control

support-agentIts own identity, not a shared key

CredentialsManaged
OwnerAna, Support
AccessScoped to 5 tools
Sign-inSSO

Identity management

Each agent gets its own identity and credentials, managed for you

ApprovedRefund of $80 for order 4790

support-agent needs youRefund $240 for order 4821. It's over the $200 limit, everything else checks out.

ApproveDeny

Human approvals

Sensitive steps wait for a person, one click to approve

Outcome · a brief sales can useMet
Attempt 1

Two claims had no source

Retry
Attempt 2

Missed one competitor

Retry
Attempt 3

Every criterion met

Pass

Outcome self-evaluation

Set success criteria. The agent checks its own work and iterates until it gets there

Recovered on its own

Shopify API timed out

Retried with backoff, order found

Context compacted, kept what matters

Ticket resolved

Recovers from errors

The harness manages context and works around failures for you

Its own cloud computer

Every agent gets its own isolated cloud machine for its tools and code. You define the agent, we run it

Cloud browserYour Chrome
Salesforceacme.my.salesforce.com/opportunities

OpportunityStageAmount

Northwind LabsProposal$84K

Contoso RetailNegotiation$61K

Fabrikam IncDemo$58K

Tailspin ToysQualified$42K

SPIRITT

Cloud and local browser

Your agents browse in the cloud, or in your own Chrome with your logins

Bills, desktop app
Entering 38 of 212 billsSave and close
SPIRITT

Computer use

Clicks and types through any app, just like a person would

scoring/route.tsscore.test.ts
1export async function POST(req) {
2  const { accounts } = await req.json();
3  const scored = await score(accounts, {
4    model: "best-combo",
5    reasons: true,
6  });
7  return Response.json(scored);
8}
42 tests passedDeployed to sales.acme.app

Frontier coding agent

Writes, tests and ships production code, in any language

Model selector

Every model that exists

Claude, GPT, Gemini, Grok, Kimi, DeepSeek, Qwen and more. Switch models without rebuilding your agent

CODINGREASONINGSPEEDCOSTCONTEXTTOOLS

Model composition

Blends the best of several models inside one agent, step by step

Your subscriptions

Claude MaxUsed for coding

ChatGPT ProUsed for research

SuperGrokConnect in one click

$0 extra for these models

Bring your AI subscriptions

Connect the Claude, ChatGPT and Grok plans you already pay for

Included in SPIRITTPay per use
apollo

ApolloPeople search

$0.02
hunter

HunterEmail finder

$0.01
lusha

LushaPhone numbers

$0.04
zoominfo

ZoomInfoCompany data

$0.05

80+ subscriptions included

Apollo, Hunter, Lusha, ZoomInfo and more for your agents, paid per use with no extra subscriptions

Slack#ops-requests
E
Emily Carter11:55
Can someone pull last month's churned accounts and draft win-back emails? @SPIRITT
👀 1
SPIRITT
SPIRITT11:58
Done. 38 churned accounts, a personal draft for each is waiting in your Gmail.
Summary: Win-back · October

Works where you talk

Tag @SPIRITT in Slack, Teams, iMessage, WhatsApp and email. It reads the thread and gets to work

#launch-planEJR
E

EmilyLet's push the launch ad budget

J

JakeCap it at $500 a day?

SPIRITTDone. Budget set for 3 channels, report every morning

Ryan is typing

Multiplayer mode

Your whole team in one thread with SPIRITT, everyone sees the work as it happens

Draft a follow-up to Sarah and move the demo to Thursday

Dictation and live conversation

Talk instead of type, or have a live voice conversation with SPIRITT

+1,500

1500+ integrations

1500+ ready to go, plus custom integrations for anything with an API

Sarah Lin, Northwind LabsOn call · 02:14

Hi Sarah, it's SPIRITT from Acme, calling about your renewal.

Sure, send me the new terms.

Call center

Buys a number, then makes and takes calls on your behalf

Voice and music
A
Ava, warm narrator
Launch theme, upbeat synth
English · Spanish · JapaneseExport

Voice and music

Voiceovers, narration, podcasts and original music, in any voice and language

this changed my morning routine

TikTok9:16

Meta4:5

YouTube16:9

Video generation

UGC, ads and explainers, always on the frontier model

Serum on travertine, hard sun, palm shadows

Image generation

Ads, product shots and brand assets on demand

Reference
Ship faster
Your version
Glow, naturally
Style matched · 3 variants

Clone videos and sites

Point at any site or video you love and get your own version, in your brand

Mounting bracket · v3Printable
48 mm

3D and CAD

Parts, products and scenes modeled from a sentence, ready to print or render

lume.com
Glow, naturallyShop the serum
sales.acme.app
Score 340

Apps, sites and games

From idea to production, full stack

PDFContract.pdf

XLSQ3 actuals

MP4Sales call

PNGReceipt scan

Extracted

CustomerNorthwind

Total$46,200

RenewsMar 2027

Uplift7%

Processes any file

PDFs, sheets, scans, audio and video, turned into answers

Works with your laptop closed

Progress and outputs live in the cloud, so a dropped connection never costs a run

Hidden revenue huntDay 23
Week 1Mapped 4 sources
Week 2$640K found
Week 3$1.24M found

Runs 24/7

Long-running sessions for weeks and months, with progress reports along the way

BillingContractsUsageCRM
2,400 agents working

Agent swarms

Agents direct other agents to split big work in parallel, hundreds or thousands at once

NoticedChurn risk is up in EMEA this week

Noticed something worth doing23 trial accounts stall at onboarding step 3. I set up an agent that sends each one a short walkthrough.

Start agentNot now

Proactive agents

Spots work worth doing and sets up agents for it

Routines and workflows

Say it once in your own words. SPIRITT runs it on schedule and keeps it healthy

Unblocked itself

Needs to read 1,200 scans

No tool exists, so it built one

invoice-ocr passed 20 tests

Processing 1,184 of 1,200

On-demand tool creation

Builds the tool it needs and unblocks itself

Certificate of IncorporationAcme Labs, Inc.Filed

EIN issued

Sales tax filed

1099s sent

Handles the paperwork

Incorporation, tax filings, permits and the rest of the bureaucracy

SPIRITT Agent•••• 4821
figmaPaid Figma$49.00
Within $500 monthly limit

Agent credit card

SPIRITT pays online with its own card, within limits you set

sales.acme.appLive
Next.js app3 regions
Postgres12 GB
DomainSSL on
File storage48 GB
Background jobs4 running
Deployed v42 in 38 seconds

Deployments and cloud resources

Domains, databases, storage and servers, handled for you

Memory
Person

Emily wants the digest on Monday mornings3 weeks ago

Decision

Q4 ads budget capped at $500 a dayyesterday

Preference

Brand voice: warm, short, no jargon2 months ago

Short and long-term memory

Remembers your people, preferences and decisions

Budgets · OctoberAlerts at 90%

Sales$4,200 of $5,000

Marketing$2,900 of $4,000

Engineering$6,100 of $8,000

Enterprise budget control

Budgets by team, project and agent, with alerts before you hit them

DesignAppSlideSheetVideo3D
Q3 Board updateRevenue up 38%

A dynamic editor for your work

The interface shapes itself around YOUR work

From agent idea to production,
SPIRITT runs the hard part.

01

Define

Describe your agent's tasks, tools and guardrails, and pick any model or a mixture. No infrastructure to wire, no agent loop to rebuild for every model upgrade

Explore the API ›
A support agent's page in SPIRITT: models, daily budget, tools and who can use it
You
Create a support agent that answers tickets from Zendesk, refunds under $200 and escalates the rest

support-agentDefined

5 tools2 modelsRefunds under $200
A weekly health report from an agent SPIRITT keeps improving
Program search · support-agent

Tuned against your criteriaCustom harness

ToolsGeneric setFitted to your stack

PromptsDefaultTuned

ContextEverythingWhat matters

Model per stepOne modelMixture

02

Tune

Set your outcomes. Program search explores the harness, tools and prompts against them and hands you a custom harness fitted to your workload, not a one-size-fits-all loop

Talk to us ›
03

Deploy

Ship it behind a clean API on managed infrastructure. Sandboxing, long-running sessions, multi-agent coordination and scoped permissions come built in, so your team works on the product, not the plumbing

Get started ›
Agents running live in the SPIRITT workspace
Your app POST /agents/support-agent

Liveresolved

SandboxedLong-runningScopedTraced
Every agent run, traced and healthy
Trace · support-agent

read_ticket0.8s

lookup_order1.1s

issue_refundApproved

reply_customerSent

Resolved · every step traced
04

Observe

Every run, tool call and decision is traced, so you always know what your agents did and why. Your agents keep getting better, and you stay in control

Explore Enterprise ›

Any integration you can imagine,
and any custom one you need

slackAnswer requests in the thread
gmailTriage and reply to the inbox
hubspotQualify and enrich every lead
stripeHandle refunds within policy
notionKeep the knowledge base current
googlesheetsReconcile and report
linearPick up and close issues
githubReview and fix pull requests
And 1500+ integrations and connectorsExplore integrations

Agents
teams deploy

Ava · Ops worker
A

AvaAI ops worker

MemoryVoice

Jake@Ava send this month's invoices?

Remembers Jake wants PDFs

Pulls the invoices from Stripe

Emails them to finance

Done, all sent
Sales agent
R

Ryan ColeVP Operations · Acme Logistics

Good fit
ResearchedWebsiteLinkedInCRM

Ryan, an idea for your dispatch team

Saw Acme opened a new hub in Denver. Teams like yours use us to…

Booked Thu 10 AMCRM updated
Support line

Incoming callEmily Carter · answered by AI

EmilyCan I move my appointment to Friday?

AgentFriday at 2 PM is open. You're booked.

Rebooked for Fri 2:00 PMBooked
Refund to Jake, with contextEscalated
Agent · Support triage
Tools

Gmail

Zendesk

Stripe

Guardrail

Refunds need a human

Run trace

Read ticket0.4s

Look up order1.1s

Draft refund0.8s

Ask EmilyApprove

GTM machine
Leads›Enriched›Outreach›Follow-ups

EEmily ParkHead of Growth · LumenMeeting

JJake MillerCOO · FieldworkReplied

RRyan BrooksFounder · TidepoolFollow-up 2

MMegan LeeVP Sales · OrbitEmailed

PipelineLive
Revenue discovery
ScanningCRMCallsSupport
Missed renewal

Northwind contract ended, no renewal sent

EEmilyOwner
Underpriced deal

Bluebird renewing below list price

JJakeOwner
Call assistant

Calling Harbor SupplyFrom your SPIRITT number

0:48

TaskConfirm Tuesday delivery and dock time

Call summarySent to you

Delivery confirmed for Tuesday

Driver calls 30 min before arrival

Logged in your CRM

Cold outreach
Email 1›Follow-up›Call›Book

EEmily“Loved your post on pricing…”Replied

JJake“Congrats on the Austin office…”Opened

RRyan“Saw you're hiring SDRs…”Booked

EmilySounds good. Does Thursday work?
Find more ideas and use cases

We run SPIRITT on SPIRITT

Our own agents run on the same managed infrastructure you get, in production, at scale. We don't sell you anything we don't use ourselves.

Our mission

Govern every agent
with fine-grained
controls

SOC 2
Type II
SAML & OIDC
SSO
Zero data
retention
An isolated sandbox for every agentAllow, ask, or deny every tool callHuman approvals for sensitive stepsScoped permissions and identitiesExecution tracing and audit logsProtected secrets and encrypted storage

Production agents
in days, not months

Buy from builders who use what they sellBuilt usingSPIRITT