Kong wants to host the agents it already governs. Kong Volcano, announced September 30, 2026 and now being pushed again with a feature called LavaFlow, is an "agent cloud": a single platform where a coding agent can create a database, add sign-in, run durable workflows and deploy an app. Kong's pitch for LavaFlow is that it makes applications and agents self-healing. The product is real and installable. The self-healing claim has no published method behind it yet.
This post separates what Kong documents from what it claims, and lists what to check before building on it.
TL;DR: Volcano in one table
| Question | Answer |
|---|---|
| What is it? | A cloud for AI agents and web apps, built to be driven by coding agents. |
| Who makes it? | Kong Inc., an API and AI gateway company. |
| What is in it? | Durable workflows, branchable Postgres, edge functions, frontend hosting, auth, realtime, file storage, approvals, sandboxes. |
| Which agents? | Claude Code, Codex, Cursor, OpenCode, Pi. |
| How to start? | npm i -g @volcano.dev/cli then volcano connect --signup. |
| What is LavaFlow? | "Self-healing applications and agents," per the Volcano site. |
| Is self-healing documented? | No method, docs or example published. |
| Pricing and hosting region? | Not published, per Futurum. |
What Kong documents
Kong's launch release calls Volcano "a new AI-native platform designed to give developers the infrastructure they need to build, deploy and operate AI agents and modern web applications." It lists these features:
- Durable workflows and jobs with leader election. The Volcano site says multi-step workflows checkpoint every step and resume where they stopped, for up to a year.
- Branchable PostgreSQL with backups and vector support. The site says a copy can be branched in under a minute, with pgvector and row-level security.
- Edge functions with autoscaling and geofencing.
- Frontend hosting for Next.js through a CDN with DDoS protection.
- Authentication and SSO with unlimited users, third-party SSO and anonymous sign-ups.
- Realtime services for database events, broadcast and presence, useful for multi-agent coordination.
- File storage on a global CDN with access policies. The site says buckets are private by default.
- Locks, queues and human approvals for multi-agent crews, with approvals that pause a run until a person signs off.
- Sandboxes: isolated environments with files, a shell and exec for running code an agent wrote. Kong said at launch that sandbox compute would follow shortly.
CTO Marco Palladino framed the pitch this way: "Developers shouldn't have to stitch together infrastructure to turn an AI agent into a production application." Kong also says Volcano is compatible with its Konnect platform for governance at scale.
The release states plainly that this Volcano is separate from the VolcanoSDK announced in 2025, which is a different, open-source TypeScript project.
What is LavaFlow, and is it really self-healing?
The Volcano site introduces LavaFlow: "With Volcano LavaFlow, you can build and run in the AI era with a self-healing platform that requires no IDEs and no human intervention to improve software autonomously."
That is a strong sentence. It claims that software improves itself without a human. The mechanism would be a technical post, documentation page, benchmark or example showing what failure LavaFlow detects, what it changes, and how it avoids breaking things. None is linked from the Volcano site, and the September launch release does not mention LavaFlow at all. What the documentation does support are the neighbors of the idea: durable functions that resume after a crash, distributed locks with renewing leases, and branchable databases that make a rollback cheap.
| Claim | What backs it up | Status |
|---|---|---|
| Workflows resume after a stop, for up to a year | Volcano site description of durable functions | Documented feature, no public test |
| Branch a database in under a minute | Volcano site | Vendor figure |
| Agents get isolated sandboxes | Release and site | Sandboxes were still to follow at launch |
| LavaFlow makes software self-healing | One sentence on the site | No method, example or benchmark found |
The practical reading: durable workflows reduce the damage when something fails, which is useful. Autonomous improvement of production software is a different and much larger claim.
How it works in practice
The Volcano site shows a prompt-first setup. You paste a setup prompt into your coding agent, which runs the CLI and opens a browser to create your account. After that, the agent can provision a database, deploy functions and add sign-in through the CLI, using skills Volcano installs for each agent it finds.
Three recipes are listed on the site: deploy the Swagger Petstore API on Postgres, host an MCP server as a Volcano function, and receive GitHub webhooks and summarize each pull request with a durable function. The MCP recipe is notable because hosting tools is a common first need. Our guide to building an MCP server covers the protocol side.
A reasonable first test is the webhook recipe. It exercises durable functions and an external API call, and you can kill a worker mid-run to see whether it resumes. That is a direct check of the resilience claim.
What analysts flag
Futurum's analysis of Volcano is the most detailed outside read. It credits Kong for naming the right problem and for covering most of what an agent app needs, then lists concerns:
- Observability gap. None of the eight services in the release provides production observability, and Kong's advanced AI observability product was not generally available.
- Governance maturity. Futurum asks whether Kong's Token Vault, AI Cost Management and Advanced AI Observability become generally available.
- Approvals. Futurum cites its own survey finding 32% of organizations have human approval gates for irreversible agent actions. Kong has not said whether approvals are on by default.
- Neutrality. The article's title says Kong now hosts the agents it governs, which sits uneasily with Konnect's pitch as a neutral governance layer across clouds.
- Unknowns. Hosting location, pricing, open interfaces for workflows and queues, and whether traces export in OpenTelemetry format.
Our suggestion, based on those gaps: prototype one agent with its logic separated from Volcano-specific calls, decide which actions need human approval, and get written answers on observability, hosting and migration before production.
Where Volcano sits among agent platforms
The agent-infrastructure space is crowded. For comparison, see our coverage of AWS Bedrock AgentCore identity and private CA, the Maritime managed cloud for production agents, the Vercel Board agent message board and Tencent's CubeSandbox for agent sandboxes.
| Platform | Strength | Question to ask |
|---|---|---|
| Kong Volcano | One CLI driven by coding agents, Postgres branching, durable functions | Hosting region, pricing, lock-in, observability |
| Hyperscaler agent runtimes | Mature identity and networking | Complexity and cost of stitching |
| Managed agent clouds from startups | Focus on production agents | Longevity and support |
We have not benchmarked Volcano against these, so the table lists questions, not scores.
What this means for builders
If you build with coding agents, the draw is speed to a working app with a real database and auth. The risks are what happens when an agent holds those keys. Give each agent the narrowest credentials, require approvals for destructive steps, and keep a way to move your data out. Postgres data can move to any Postgres host. Application code written against Volcano's workflow and realtime APIs may not.
If you run a team, treat self-healing as a roadmap item until Kong shows a method. Ask for a written description of what LavaFlow changes on its own, how changes are reviewed, and how to turn it off.
Questions people ask about Kong Volcano
Is it open source? Kong has not said so. The CLI is distributed on npm and Homebrew, but the platform itself is a hosted service, so assume it is not self-hostable until Kong says otherwise.
Does it replace Supabase or Vercel? It overlaps with both: branchable Postgres, auth, storage and Next.js hosting. The difference Kong stresses is the agent focus, with durable workflows, approvals and sandboxes in the same product. We have not run a side-by-side comparison, so judge by your own workload.
Is it safe to let an agent deploy to it? That depends on what you grant the agent. Volcano offers approvals that pause a run until a person signs off, but Kong has not said whether they are on by default. Turn them on for deletes, schema changes and anything that touches user data.
Is the new Volcano the same as the 2025 Volcano SDK? No. Kong says they are separate products, and some write-ups mix them up.
How to verify before you adopt
- Install the CLI in a throwaway project and run the webhook recipe.
- Kill a worker mid-workflow and confirm it resumes from the last checkpoint.
- Branch the database and confirm the branch is isolated.
- Ask Kong for pricing, hosting region and an export path in writing.
- Check whether traces can reach your existing observability tools.
Bottom line
Volcano is a coherent bundle for people who let coding agents build full apps, and the durable-workflow and branching features are concrete. The self-healing label is a marketing sentence until Kong publishes how it works. Try it on a small project and judge it by what you can reproduce.
Details reflect Kong's launch release, the Volcano site and Futurum's analysis as of November 19, 2026. Features and pricing may change.
