Turnkey platform for serious teams. No lock-in.
Your agents write the code. OpenFlame ships it and keeps it running.
Describe the system you want. OpenFlame creates a Kubernetes cluster in your own cloud, sets up the full platform, and gives every developer a workspace where AI agents build services, web apps and mobile apps inside a structure that holds up in production. Disconnect OpenFlame at any time and everything keeps running in your account.
A booking app for dog groomers: customer mobile app, salon web dashboard, payments, SMS reminders.
- Cloud and cluster size done
- Create the cluster done
- Install the platform done
- Start your workspace done
- Link your AI accounts done
- Choose a development model done
- Plan and build the first version building
Why structure matters
Agents are fast. Production needs rules they can follow.
An agent with no constraints produces a different architecture every time. OpenFlame gives agents a paved path, one opinionated project: where services live, how data is defined, how APIs are generated, how releases move. That is what makes the output reviewable, deployable and operable.
A sentence becomes a plan
The Architect drafts data models, APIs, services, UIs, sign-in and integrations as cards you can edit before any code exists.
The plan becomes a pull request
The platform's generators write the scaffolding. A coding agent then builds the features in your workspace.
Every PR gets an environment
Its own namespace, a clone of the database, an HTTPS address, and automatic teardown.
Releases move by review
Development to staging to production, with contract checks, canary analysis and rollback.
Develop
A cloud workspace for every developer, with their own AI agents inside
Each developer gets a workspace on the development cluster, next to the real database, cache, message bus and services. Nothing to install on a laptop.
Browser IDE with every toolchain
VS Code in the browser with Go, Python, Node, Rust, Java and .NET ready, plus kubectl, the GitHub CLI and a Postgres client.
Claude Code, Codex, Cursor and opencode
The coding agents are installed and configured. Use the terminal, or drive them from the console.
Bring your own AI subscriptions
Connect your Claude, ChatGPT, Google, Grok, Kimi, Devin or Meta plan. Credentials stay inside your workspace, and OpenFlame adds no charge for model usage.
Describe a change, get a pull request
Each workstream runs on its own branch. Watch the agent work, read the diff, preview the UI or run the service, then open the PR. Works from a phone.
Agents start with the facts
A governed project record of architecture, intent and decisions is validated and handed to every agent run, with a playbook and skills for the platform.
Agents that work until a check passes
Goal-driven runs with limits on cost, time and iterations, human approvals, and templates such as PR Shepherd, Flake Hunter and Coverage Ratchet.
Attach a real debugger
Breakpoints on running services in all six languages, web apps in a browser inside the workspace, and React Native on a device.
Test your branch in the shared cluster
Run a service from your workspace and only your requests reach it. Everyone else keeps using the deployed version.
Captured email on development
Mail your app sends on a development cluster lands in an inbox in the console, never in a customer's.
Data model
Define a type once. The database and every API follow.
A data model is a JSON Schema file in your repository. Saving it regenerates everything that depends on it, so services, UIs and agents always agree on the shape of your data.
# data/schema/appointment.json { "title": "Appointment", "properties": { "id": { "type": "string" }, "salon": { "$ref": "salon.json" }, "starts_at":{ "format": "date-time" }, "status": { "enum": ["booked", "done"] } }, "x-openflame": { "persist": true, "hasura": true } }
- SQL migrationsUp and down pairs, with a preview. References become foreign keys.
- GraphQLTables tracked in Hasura with relationships and permissions.
- ProtobufMessages with field numbers that never change or get reused.
- OpenAPIOne document per API, with your types as components.
- gRPC and RESTService stubs in six languages and a JSON gateway in front.
- Edge rulesPublic and signed-in routes enforced at the gateway.
- TypeScript SDKsTyped clients for your web and mobile apps.
Services
Service-oriented from the first commit
Generate a service in the language that fits the job. Every one gets the same shape: gRPC inside, REST at the edge, health checks, tracing, and the caller's identity from a signed token.
- Go
- Python
- Node (TypeScript)
- Rust
- Java
- C#
- Declare what a service needs. Database, cache, storage and message bus connections are a line in a file, wired in as environment variables.
- Set resources per environment. Replicas, autoscaling and rollout strategy, with every value traced to where it was set.
- Events and schedules. Trigger a service or a workflow from a webhook, a cron schedule, a storage upload or a message.
- A CI check keeps the structure. Services that drift from their template fail validation before they merge.
Web and mobile
React on the web. Expo on the phone.
New projects start with a working web app and a working mobile app, both with sign-in, typed API clients and GraphQL already connected.
- Live preview per developer. Web in the browser, mobile through Expo Go with a QR code.
- Connect Expo once. Store-signed iOS and Android builds run through EAS from your pipeline. Android development builds run in your cluster.
- Over-the-air updates follow releases. Each environment has its own update channel, moved by promotion.
- One logo, every icon. Upload a logo and get favicons, app icons and splash screens committed to the repository.
- Firewatch watches the Expo SDK. When an SDK upgrade fixes a vulnerability, it proposes the upgrade with a breaking-change checklist.
Ship
CI/CD that runs in your cluster and reports to your pull request
Argo Workflows builds, Argo CD deploys, and git is the source of truth. There is no external CI service to configure and no deploy credential outside your cloud.
Builds only what changed
Change detection across the monorepo, rootless BuildKit with a layer cache, and a CI node pool that scales from zero.
Preview per pull request
The PR's own services and UIs, a cloned database, seed data and an HTTPS address. Idle previews are removed.
Review and test steps you choose
Add AI code review and end-to-end test providers to the pipeline, aimed at the preview address.
Promotion with contract checks
A release plan verifies that migrations, GraphQL, APIs, services and apps are compatible before anything moves.
Canary and blue-green
Traffic shifts in steps while error rate and latency are measured. A failed analysis stops the rollout.
One grid for every setting
See each value per environment and cluster, where it came from, and where environments have drifted apart.
Platform services
What you would expect from a cloud provider, without the lock-in
Each one is installed, monitored, backed by generated credentials and browsable from the console. All of them are open source.
PostgreSQL
CloudNativePG with replicas, continuous backup, point-in-time restore, clones and pgvector. Or a managed Cloud SQL or RDS instance.
Valkey
Sentinel for failover and a proxy for clients that need a single address.
NATS
Publish and subscribe for services, and JetStream behind the event system.
RustFS
S3-compatible storage in the cluster. Point any purpose at an external bucket without changing application code.
Hasura
Queries, subscriptions and REST endpoints over your data model, behind signed tokens.
Haraka
Outbound mail relayed through Cloudflare with signing handled for you. Inbound mail arrives as messages on the bus.
Ory Kratos
Password, passwordless, social and two-factor sign-in, with flows and email templates edited in the console.
Infisical
One project per application and environment, synced into the cluster, with reveal logging and credential rotation.
Model serving
Optional: deploy open models from Hugging Face on your own GPUs with an OpenAI-compatible endpoint.
Integrations
Call outside services through a gateway that holds the keys
Your services never see a third-party credential. They call the integration gateway, which adds the key and applies the policy you set.
- Rate limits and quotas. Requests per second, calls per day or month, per integration.
- Dollar budgets for model providers. Token usage is priced and capped.
- Caller allow-lists. Decide which services, loops and tools may use each integration.
- Metering and alerts. Usage per caller in Grafana, with an alert at 80% of quota.
- Inbound webhooks. Signature, token or header checks per endpoint before a request reaches your service.
Operate
When an alert fires, the diagnosis is already written
Every service is instrumented from the template. Alerts open incidents, and an agent investigates each one with the same tools an engineer would use.
Metrics, logs and traces
Grafana, Mimir, Loki, Tempo and Alloy, with OpenTelemetry wired into services and dashboards for APIs, databases and queues.
Incidents diagnosed by an agent
It reads the runbook, logs, metrics, pod state and your project record, then posts a diagnosis and a recommendation.
People approve the fix
Remediation runs as a loop that a person starts, with approvals, a budget and a full record of each step.
Written once, used every time
Runbooks are matched to alerts by label and handed to the agent as its starting point.
Service level objectives
Objectives measured from mesh traffic, and a health page for the whole project.
Observability inside the workspace
Coding agents can query your dashboards, logs and metrics directly while they work.
Security
Hidden behind Cloudflare, watched by Firewatch
Security is part of how a cluster is created, how code is built and how fixes arrive. It is on from the first deploy.
The edge
- No public load balancer for web traffic. Each cluster connects outward to Cloudflare through its own tunnel. DNS, certificates and routes are created automatically.
- No kubeconfig leaves your cloud. An agent in the cluster dials out to OpenFlame with least-privilege permissions.
- Admin tools behind sign-in. Console, dashboards and data browsers admit only members of your organisation.
- Services accept known callers. A mesh policy on every service restricts who can reach it.
- An audit log. Deletes, secret changes, approvals and membership changes are recorded, and secret reveals are logged.
Firewatch
- Scans what you run. Dependencies, base images, the images your infrastructure deploys, misconfigurations, secrets in git history, licences, end-of-life runtimes and the live cluster.
- Ranks by real risk. Known exploitation, exploit probability, live incidents, and whether your code calls the vulnerable function.
- Opens the fix. Renovate handles updates, an AI agent handles what Renovate cannot, and patched images are rebuilt from fixed packages.
- Never merges. Every change waits for a person, and each PR is driven until your own checks pass.
- Handles a leaked key. A runbook for revoking and rotating, with direct links for 20 issuers.
Firewatch is open source under Apache-2.0 and also runs on its own.
Your cloud, open source
Runs in your account. Built from parts you can inspect.
OpenFlame creates clusters in your cloud account with your credentials, or installs onto a cluster you already have. Your code, configuration and data stay in your repository and your cloud.
- Google GKE
- Amazon EKS
- Azure AKS
- DigitalOcean
- Oracle OKE
- Scaleway
- Vultr
- Your own cluster
| Layer | What you get | Built on |
|---|---|---|
| Provisioning | Clusters and managed databases in your account | Crossplane 2.4 |
| Delivery | Builds, deploys, events and progressive rollouts | Argo CD · Workflows · Events · Rollouts · BuildKit |
| Database | PostgreSQL with backup, restore, clones and vectors | CloudNativePG · PostgreSQL 16 · pgvector |
| Cache | Key-value cache with failover | Valkey 9 · Sentinel |
| Messaging | Message bus and event streams | NATS 2.15 · JetStream |
| Storage | S3-compatible object storage | RustFS 1.0 |
| API | GraphQL, REST and gRPC behind one gateway | Hasura 2.51 · Istio · grpc-gateway |
| Identity | Sign-in and access rules at the edge | Ory Kratos · Ory Oathkeeper |
| Secrets | Secret storage, sync and rotation | Infisical · External Secrets |
| Observability | Metrics, logs, traces, alerts | Grafana · Mimir · Loki · Tempo · Alloy |
| Edge | Tunnel, certificates, DNS and mail relay | cloudflared · cert-manager · Haraka 3.4 |
| Security | Scanning, updates and patched images | Firewatch · Trivy · osv-scanner · Renovate |
| Models | Optional model serving on your GPUs | KServe · vLLM |
Pricing
A flat platform fee. Your cloud bill and your AI plans stay yours.
OpenFlame does not mark up compute or model usage. Build for free, and pay when you add staging or production.
Launch
Build a first version.
- 1 development cluster
- 2 developers
- Previews per pull request
- Firewatch scanning
- Community support
Team
Ship to production.
- 3 clusters
- 5 developers
- Promotion and rollouts
- Incident diagnosis
- Firewatch fix loop
- Email support
Business
Several environments or clients.
- 8 clusters
- 20 developers
- Audit log export
- Scheduled upgrade windows
- Priority support
Enterprise
Regulated and sovereign teams.
- Run OpenFlame itself in your account
- Air-gapped installs
- Named support
- Custom terms
Extra developers are $39 a month and extra clusters $199 a month. Prices are in US dollars and exclude what you pay your cloud and AI providers.
Made for the whole team
One platform from API to UI to on-call
Typed clients and live previews
Generated SDKs, GraphQL, working sign-in, a preview per branch, and mobile builds without a build machine.
Services that fit together
Shared types across languages, generated APIs, a real database and bus in development, and breakpoints on live requests.
A cluster you can reason about
Everything declared in git, releases by review, alerts with a diagnosis attached, and fixes that arrive as pull requests.
Questions
Before you ask
Do I need to know Kubernetes?
No. You describe services, data and environments in the console or in plain files, and every change is a pull request you can read. Kubernetes is there when you want it, with a resource browser, logs and events.
Where does my code and data live?
In your git repository and your cloud account. You own the clusters, including their control planes. OpenFlame holds project metadata and coordinates work, and never holds a kubeconfig for your cluster. The image registry and the secret store are moving onto your own build cluster, so those will be yours too.
Which AI models can I use?
The ones you already pay for. Developers connect their own subscriptions inside their workspace. Platform agents call a model provider you configure, through the integration gateway, with a budget you set.
Will an agent change production on its own?
No. Agents propose and people approve. Incident response produces a diagnosis and a recommendation. Firewatch opens pull requests and never merges them. Coding agents cannot run against staging or production clusters.
What if I want to leave?
Everything OpenFlame installs is open source and everything it generates is in your repository. You own the clusters and their control planes, so they keep running without OpenFlame.
Which git providers work?
GitHub today, for builds, previews and checks.
Early access
Bring a system you want built. Leave with it running in your cloud.
OpenFlame is onboarding design partners: teams building a multi-service product with AI agents, and studios that build for clients. Tell us what you are building and which cloud you use.
hello@openfla.me