Velt
Also known as: Velt, Inc., Velt SDK, Velt review agents, velt.dev
Velt is an embeddable SDK that adds review and approval to AI native apps, so people can check what agents produce before it ships. Review agents read a page, dataset or chart and post findings as comments, agent suggestions wait for a person to accept them, and a workflow engine routes work between agents and human approvers with quorums, deadlines and escalation. A memory layer records each decision so later reviews follow precedent.
Velt, Inc., a Y Combinator company, sells an SDK that software teams embed to add comments, presence, multiplayer editing, recording, notifications and, increasingly, review and approval for work that AI agents produce. Its pitch is embeddable review and approval for AI native apps. OpenEnvoy runs more than 500,000 reviews in production with more than 2 million review decisions, X shipped full collaboration features in under a week, Leadpages describes three people's worth of work shipped in weeks, and MetaImpact raised weekly active users by 26%. Logos include Google, Varonis, Freshworks, Pendo, HeyGen and Runway.
Review agents are configurable AI reviewers. A team gives one instructions and a strategy, then starts a run against a URL: the agent gathers text, screenshots, HTML, CSS, links, the accessibility tree and Lighthouse results, can crawl several pages or emulate mobile and desktop devices, and reads data from REST endpoints or from charts and dashboards exposed through MCP.
Velt handles context gathering, the model call, deduplication and sanitizing, then posts each finding as a comment with its severity, category, suggested fix, selector and confidence. Agent suggestions become reviewable edits that a person accepts or rejects, and approved changes flow back to the app through a webhook with a record of who approved them.
The Review Workflow Builder, in beta, describes a review process once as a graph of agent, human, notification and webhook steps. Groups run steps in parallel with an approval quorum, a rejection can loop back for another attempt, and a step that misses its deadline escalates. Runs start from the API, an inbound webhook from GitHub, Vercel or a custom source, a cron schedule or an installed app, and every state change arrives as a signed, sequenced event. Memory, also in beta, records each decision with its reasoning and who made it, derives reviewer patterns and answers questions about precedent over a knowledge base built from uploaded guidelines.
The free Hacker plan covers development with 100 monthly active documents, while Growth and Enterprise are priced by contract. Enterprise adds partial or full self hosting, multi region hosting, isolated storage, SOC 2 Type II and HIPAA with a business associate agreement, and SAML single sign on with Okta user sync is available. Velt suits product teams building AI features into document, design, analytics or workflow apps that need a human check before agent output goes live, while teams wanting a finished standalone review tool rather than components to embed will find an SDK.
Vendor details
Canonical URL
https://velt.dev/
Category
Agent infrastructure
Subcategory
Embeddable review, approval and review agents for AI native apps
Company status
independent
Use cases & customers
Primary use cases
Target customers
Deployment options
Integrations
Embeddable React and JavaScript components, REST APIs and Node and Python SDKs, webhooks, an MCP server for Velt data, review agents that read URLs, REST endpoints and MCP tools, and workflow triggers from GitHub, Vercel, cron and custom webhooks.
In practice
A procurement platform's agent drafts invoice corrections as suggestions, and a finance reviewer approves each one before the webhook applies it.
A review agent checks a marketing page for broken links and accessibility issues and leaves findings as comments for the designer.
A pull request on GitHub starts a workflow where an agent reviews the preview, two approvers must agree and a missed deadline escalates.
Sources & related URLs
Related / legacy domains
Agentic Index coverage score
11.5 / 14 capabilities · 82%
| Integrations & Tool Calling | Full |
|---|---|
|
Review agents pull data from any REST endpoint and from charts or dashboards exposed through an MCP server, then write findings back as comment annotations through Velt's APIs. Workflow steps send email or Slack notifications and call a team's HTTPS endpoint, and once a person approves an agent's suggestion, the change is applied in the customer's own system through a signed webhook. Sourcevelt.dev/docs/ai/agents/overview.mdread 2026-11-20 |
|
| Workflow Orchestration | Full |
|
The Review Workflow Builder, in beta, runs a review process defined as a graph of agent, human, notification and webhook steps with up to 100 nodes. Groups run steps in parallel under a quorum such as everyone or any two approvers, a rejection can loop back for a bounded retry, and agent steps can hold the workflow until findings are ready. The engine validates each graph when it is saved. Sourcevelt.dev/docs/ai/approval-engine/overview.mdread 2026-11-20 |
|
| Knowledge Grounding & RAG | Full |
|
Memory, in beta, turns uploaded guideline files in PDF, CSV, Excel or text into markdown, extracts their rules and embeds them in a workspace knowledge base. Its ask endpoint returns a cited answer with a confidence score and its suggest endpoint recommends a decision for a new item. Review agents also gather fresh page content, accessibility data and Lighthouse results for each run. Sourcevelt.dev/docs/ai/memory/overview.mdread 2026-11-20 |
|
| Human Oversight & Guardrails | Full |
|
Human review is the product. Agent proposals arrive as comments or suggested edits that a person must accept before anything changes, with a record of who approved each one, and workflow human steps wait for approve or reject decisions, optionally under a quorum or from named people. Steps that miss their deadline escalate, and review agents sanitize and deduplicate findings before posting. Sourcevelt.dev/docs/ai/approval-engine/overview.mdread 2026-11-20 |
|
| Security, Identity & Governance | Full |
|
Velt supports SAML single sign on and Okta user sync, issues JWT auth tokens, and offers a permissions API plus share and invite controls that decide which users reach which content. Enterprise contracts add SOC 2 Type II, HIPAA with a business associate agreement, a data processing agreement, isolated server and data storage that is never mixed with other tenants, and custom security reviews. Sourcevelt.dev/pricingread 2026-11-20 |
|
| Observability & Auditability | Full |
|
Activity logs record actions on comments, suggestions and documents and can be pulled through the API or shown in an embedded log. Workflow state changes arrive as signed events with a sequence number, so a team can replay a run from any point, and each review agent finding carries its severity, confidence and selector. Token usage analytics break down by agent and model. Sourcevelt.dev/docs/ai/agents/overview.mdread 2026-11-20 |
|
| Memory & State Persistence | Partial |
|
Every review decision, whether a person's or an agent's, is saved automatically as a judgment with the decision, its reasoning, who made it and what it applied to, and Memory derives reviewer profiles and decision patterns from them. Reads can be scoped to one organization or a set of documents. Judgments are read only, with no stated retention period or way to purge them. Sourcevelt.dev/docs/ai/memory/overview.mdread 2026-11-20 |
|
| Deployment & Data Residency | Full |
|
Velt Cloud runs in 42 regions, with data provider regions in the US, EU and Asia Pacific and residency pinning. Partial self hosting keeps comments, recordings, files and user PII in the customer's own MongoDB or PostgreSQL and S3 compatible storage while Velt holds only identifiers, and full self hosting runs on Google Cloud, with AWS and Azure in closed beta. Sourcevelt.dev/self-hostingread 2026-11-20 |
|
| Prebuilt Agents / Templates / Packs | Partial |
|
More than 15 collaboration features arrive as prebuilt, customizable components, from comments and presence to recording and huddles. Review agents start from strategies such as page review, REST data review and MCP tools, but each agent's instructions and configuration come from the customer. Workflow definitions are written as JSON, with documented patterns rather than a library of ready review processes. Sourcevelt.dev/docs/ai/agents/overview.mdread 2026-11-20 |
|
| Triggers & Channel Coverage | Full |
|
Workflow runs start from a backend dispatch call, an inbound webhook from GitHub, Vercel or a custom signed source, a cron schedule, or an installed GitHub App or Vercel integration. Review agents also start directly through their API. People act on the results inside the host app through embedded comments and notifications, so reviewers see agent findings where they already work. Sourcevelt.dev/docs/ai/approval-engine/setup.mdread 2026-11-20 |
|
| Model Flexibility & Routing | Partial |
|
Review agents using the MCP tools strategy run on Claude or Gemini, and the Rewriter runs on a major model provider, with token usage reported per agent and per model. Velt names no full model list, customer selection of the model for each agent or option to bring an API key. The customer's own agents keep whatever models they already use, since Velt reviews their output. Sourcevelt.dev/docs/ai/agents/overview.mdread 2026-11-20 |
|
| APIs / SDKs / MCP Extensibility | Full |
|
Version 2 REST APIs cover comments, activities, authentication, permissions, review agents, workflows and memory, with Node and Python backend SDKs and basic and advanced webhooks. A beta MCP server gives MCP clients access to Velt data, and separate MCP servers search the docs and install components into React and Next.js projects. Front end SDKs embed every component. Sourcevelt.dev/docs/llms.txtread 2026-11-20 |
|
| Testing, Debugging & Optimization | Partial |
|
Each finding carries a confidence score, and Memory's stats and decision patterns show how often reviewers accept or reject agent output over time. Running a review agent again replaces the previous run's suggestions on the document, so older results do not pile up. No harness scores a changed agent configuration against test cases before it reaches production reviewers. Sourcevelt.dev/docs/ai/memory/overview.mdread 2026-11-20 |
|
| Browser / Computer-use | Partial |
|
Review agents load the pages they check, capturing screenshots, HTML, CSS, links, the accessibility tree and Lighthouse results, emulating mobile or desktop devices and crawling across several URLs. That is a headless read of each page for review. The agents do not click, type or work through a site's forms or flows, and no desktop or remote computer control is part of Velt. Sourcevelt.dev/docs/ai/agents/overview.mdread 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
Free Hacker plan for development; Growth and Enterprise by contract
Monthly active documents under contract
Included quota
Hacker includes 100 monthly active documents, every feature, 20 simultaneous connections, 5,000 comments and notifications, 2GB of file storage and one team member, for development environments only.
What is public
Velt's Hacker plan is free forever for development, with 100 monthly active documents and no production deployment. Growth and Enterprise are priced on monthly active documents under contract.
Billing mechanics
Billing applies only to Velt SDK create, read, update and delete operations, measured as monthly active documents. Growth adds production use, up to 100,000 simultaneous connections, unlimited comments, 2TB of storage, REST APIs and a 99.999% uptime SLA.
Sales call required
Yes, required for paid access
Free / trial
Free Hacker plan for development
Commercial notes
Enterprise adds self hosting in the customer's database or cloud, multi region hosting, isolated storage, SOC 2 Type II, HIPAA with a business associate agreement, a dedicated success manager and implementation help. AI Enhancements are an add on, and startup and volume discounts are available.
Missing data
Velt quotes the Growth and Enterprise rate per monthly active document and the AI Enhancements add on after a demo.
Related vendors
- AgentOps — Agent observability and debugging platform: open source SDKs trace…
- AgentPhone — AgentPhone gives the AI agents a team builds their own US or…
- Agno — Python agent framework and AgentOS runtime (formerly Phidata) for…
- Airbyte — Airbyte moves data from operational systems into warehouses and…
- AIsa — Resource and payment gateway for AI agents, with one key to 110+…
- Albato — Albato is a no code integration platform that sells an embedded,…