Back to vendors
S

SafetyKit

Also known as: Safety Kit, safetykit.com

Visit site
Entry priceQuote only, after a demoFull pricing detail

SafetyKit builds AI agents and decision models for trust and safety, risk and compliance teams at marketplaces, payment companies and content platforms such as Patreon, Eventbrite, Upwork and Etsy. Its models read each platform's written policies, review content, products, users and merchants, and return labels and actions through webhooks, removing clear violations automatically and routing uncertain cases to people. More than 200 pretrained policies, a full API with SDKs and golden set testing come with the hosted service.

SafetyKit builds AI agents and decision models for trust and safety, risk and compliance teams at marketplaces, payment companies and content platforms. Customers include Patreon, Eventbrite, Upwork, Character.ai, Etsy, Discord, Lyft, Square, Faire, Depop, Grindr, Lime and Kickstarter, along with one of the four card networks. It raised $27 million from Ribbit, First Round Capital and Y Combinator, with angels including former Stripe COO Claire Hughes Johnson, and one of its cofounders built Stripe's risk platform. Upwork measured SafetyKit at 95% accuracy against 80% for human reviewers.

Platforms send users, products, transactions and messages through a Data API, along with backend events such as sign ups, uploads, product changes, user reports and moderation decisions, plus browser telemetry. SafetyKit's models read each platform's written policies directly, apply the right regional or regulatory version, and return labels, structured fields and actions through webhooks.

A content team can set an image policy to remove violations automatically at high confidence and send uncertain cases to human review, where reviewers see the evidence, guidelines and suggested actions together. Live streams get instant action on violations, and linking events that share an IP address surfaces coordinated abuse such as mass account creation or ban evasion.

The work spans content moderation across text, images, video, live streams and audio, including AI generated media; marketplace enforcement against counterfeit, recalled and hazardous products and illegal wildlife listings, where one $10 billion marketplace saw SafetyKit begin removing more than 100 violations a day; and merchant risk review for payment platforms, where agents review hundreds of pages per merchant across websites, social media and public records, assign merchant category codes with 98.8% accuracy and approve compliant merchants within seconds.

A library of more than 200 pretrained policies covers card network compliance, regulated products, platform safety and marketplace integrity, with ready policies for the Take It Down Act, the EU Digital Services Act and the UK Online Safety Act. SafetyKit also sells expert labeled datasets and decision models, which teams can call through the API or use to train their own classifiers.

Teams test policies on live and historical content before launch, track accuracy against golden sets, and send final outcomes back through an Agent Decisions API that records whether SafetyKit's review matched. SafetyKit sells through demos with no public price list and runs as a hosted service. It suits trust and safety, risk and compliance leaders at large marketplaces and payment platforms who want to move review work off outsourced teams, while teams that need a private deployment or published security attestations will find a hosted API first.

Vendor details

Canonical URL

https://www.safetykit.com/

Category

Enterprise operations agent

Subcategory

Trust and safety, risk and compliance review agents for marketplaces and payment platforms

Company status

independent

Use cases & customers

Primary use cases

Moderate text, images, video, live streams and audio against written policiesRemove counterfeit, recalled and hazardous listings on marketplacesReview and approve merchants for payment platforms

Target customers

Trust and safety teams at marketplaces and content platformsRisk and compliance teams at payment companiesPlatforms replacing outsourced human review

Deployment options

Hosted API

Integrations

Platforms send users, products, transactions, messages and backend events through the Data API, add browser telemetry through a beta SDK, and receive labels, fields and actions through webhooks, with SDKs in TypeScript, Python, Kotlin and Java.

In practice

A marketplace sends new listings through the Data API and SafetyKit removes recalled products while sending uncertain matches to reviewers.

A payments company has SafetyKit review a new merchant's website and public records, assign its merchant category code and approve it within seconds.

A creator platform moderates live streams and acts on policy violations as they happen.

Agentic Index coverage score

9.0 / 14 capabilities · 64%

Integrations & Tool Calling Partial

Platforms push users, products, transactions and messages into SafetyKit as JSON through the Data API, and SafetyKit sends labels, fields and actions back through webhooks for the platform's own systems to apply, which is how takedowns and workflows are triggered. No catalog of named connectors into case management, payment or commerce tools sits alongside that API and webhook exchange.

Sourcedocs.safetykit.com/using-data-api/core-conceptsread 2026-11-20

Workflow Orchestration Partial

Each namespace holds one object type with its own review workflow, rules, outputs and schema, and one entity can run through several namespaces at once, such as prohibited items and counterfeit risk, with success and failure events per workflow. SafetyKit runs thousands of such workflows in production, while the platform itself decides which result wins when namespaces disagree.

Sourcedocs.safetykit.com/using-data-api/core-conceptsread 2026-11-20

Knowledge Grounding & RAG Full

SafetyKit models read a platform's written policies directly, manage exceptions across rule sets and apply the right regional or regulatory version of a policy during detection. Expert labeled data on obscure brands, niche goods, recalls and new scam patterns is refreshed as they appear, so a listing can be matched to a recall notice even when it never names the product.

Sourcesafetykit.com/blog/launching-new-safety-policies-in-a-few-hours-instead-of-monthsread 2026-11-20

Human Oversight & Guardrails Full

A policy configuration sets actions by confidence, removing content automatically at high confidence and sending low confidence cases to human review. Ambiguous cases arrive with full context, and reviewers see the evidence, guidelines and suggested actions in one place, while trust and safety teams set the policies and decide edge cases before low risk decisions are automated.

Sourcesafetykit.com/content-moderation-with-airead 2026-11-20

Security, Identity & Governance Partial

Requests carry a team API key as a bearer token from the backend, and keys can be exchanged for short lived access tokens lasting five minutes to 24 hours that cannot mint further tokens and can carry a user ID and session ID for auditing. No security attestation, user roles or single sign on appear beside those API controls for the hosted service.

Sourcedocs.safetykit.com/authenticationread 2026-11-20

Observability & Auditability Partial

Each object's result arrives as a workflow success or failure event with its final labels, structured fields such as confidence and the label changes from that run, and teams can poll request status or download full result files to reconcile. A dashboard sits behind the team login, but reasoning or evidence fields do not come with the results that platforms receive.

Sourcedocs.safetykit.com/using-data-api/receiving-resultsread 2026-11-20

Memory & State Persistence Partial

SafetyKit keeps each object under a stable ID, reprocesses it when it changes and reports which labels were added or removed against its prior state. It stores a content hash beside each content ID, normalizes email addresses and links events that share an IP address to surface coordinated activity such as mass account creation or ban evasion across a platform's users.

Sourcedocs.safetykit.com/events/backend-eventsread 2026-11-20

Deployment & Data Residency Not documented

SafetyKit runs as a hosted service: platforms send data to its API from their backends, browser telemetry goes to its own ingest endpoint, results come back through webhooks and service status shows on a public status page. No choice of hosting region, private cloud deployment or self hosted option appears for teams with data residency requirements.

Sourcedocs.safetykit.comread 2026-11-20

Prebuilt Agents / Templates / Packs Full

A library of more than 200 pretrained policies covers card network compliance, regulated products, platform safety and marketplace integrity, filtered by marketplace, payment or content platform and by moderation, fraud, compliance or AI safety use. Ready policies also cover the Take It Down Act, the EU Digital Services Act and the UK Online Safety Act, and deploy in hours.

Sourcesafetykit.com/policiesread 2026-11-20

Triggers & Channel Coverage Full

Reviews start from object ingestion and from seven backend event types sent as they happen: messages and transactions between users, sign ups, account changes, uploads, product listings, user reports and moderation decisions. A beta browser SDK adds session telemetry, live video produces violation events, and moderation covers text, images, video, live streams and audio.

Sourcedocs.safetykit.com/events/backend-eventsread 2026-11-20

Model Flexibility & Routing Not documented

Platforms either call SafetyKit's own decision models through the API or train their own models on its expert labeled datasets, and SafetyKit worked with OpenAI, ROOST and Discord on the release of OpenAI's open weight safety models. No choice of underlying model provider, and no routing between models by task or cost, comes with the decision models and agents that SafetyKit runs for its customers.

Sourcesafetykit.comread 2026-11-20

APIs / SDKs / MCP Extensibility Full

A public REST API reference with dated versions covers data ingestion, batches, schema settings, request status, result downloads, events, access tokens and agent decisions, with SDKs for TypeScript, Python, Kotlin and Java that handle authentication, retries and types. Webhooks deliver results, and a beta browser SDK collects session telemetry for evaluation.

Sourcedocs.safetykit.comread 2026-11-20

Testing, Debugging & Optimization Full

Teams test new policies on live and historical content before launch, watch outcomes and refine the wording, and accuracy is tracked continuously against golden sets, with misclassifications corrected within 24 hours. An Agent Decisions API takes each final human or system outcome and reports whether SafetyKit's earlier review matched it, measuring review quality over time.

Sourcedocs.safetykit.com/agent-decisionsread 2026-11-20

Browser / Computer-use Partial

In merchant risk review, SafetyKit's agents investigate across the web, reviewing hundreds of pages per merchant and analyzing text, images, video and metadata from websites, social media and public records, and they evaluate harmful websites on content, ownership and payment flows. Operating a browser or desktop to carry out tasks inside the platform's own tools is not part of that work.

Sourcesafetykit.com/blog/scaling-merchant-risk-review-the-future-of-trust-and-safety-for-payment-platforms-2read 2026-11-20

The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded

Pricing

Quote only, after a demo

Enterprise agreements for review workflows, decision models and labeled data

What is public

SafetyKit sells its review agents, decision models and labeled datasets through demos and conversations with its team, with prices quoted by sales.

Billing mechanics

Platforms send data for review through the Data API, and each agreement sets the workflows, policies and volumes it covers.

Sales call required

Yes, required for paid access

Free / trial

No free tier or trial

Commercial notes

More than 200 pretrained policies deploy in hours, and teams can either call SafetyKit's decision models or license its labeled datasets to train their own.

Missing data

SafetyKit quotes every agreement after a demo, so volume tiers and per review rates sit in the contract.

Agentic Index verified 2026-11-20

Contact us

Found a vendor we missed? Have feedback on the index? We'd love to hear from you.