Last.Backend: a way to production for AI agents
Isolated runtimes for Claude Code, Codex, Cursor and your own agents, plus a policy-checked path from sandbox to production. Self-hosted, in your cloud, or managed.
why
Agents write code faster than anyone can read it
In the editor, an agent is an assistant. Once it touches real infrastructure, most teams can't answer these questions:
- Where does code that no human has read actually run?
- What can the agent reach on the network, and with whose tokens?
- Can it see production secrets?
- Who approves a release, and how?
- What happens when the agent hits its budget mid-task?
- Afterwards, can you tell which agent changed what, from which prompt, and who approved it?
how it works
Isolate, govern, ship
Isolation
Every agent gets its own runtime: a Firecracker or Cloud Hypervisor microVM by default, containers for trusted code.
- microVM by default
- Snapshots and forks
- Bring your own runtime
Policy as code
Permissions, budget and network live in YAML. The agent is its own principal and never borrows a developer's tokens.
- Deny-by-default egress
- Secrets injected at the proxy
- Audit log with SIEM export
A path to production
Sandbox, preview, staging, production. Production needs human approval, with canary rollouts and automatic rollback.
- Human approval
- Canary releases
- Production errors come back to the agent as tasks
pipeline · how a change ships
From sandbox to production
We have built this flow into our clients' CI/CD for years: environments, checks, a manual gate before production and a rollback plan. Last.Backend gives it to agents out of the box.
The agent gets its own isolated runtime to build, test and experiment without touching anything else.
- createA dedicated microVM or container per task
- egressNetwork is closed; allowed hosts are listed in policy
- budgetOn limit the runtime is snapshotted, not lost
Outcome: A change built and tested in isolation
runtimes
One contract, many runtimes
The control plane is separate from the runtimes, so runtimes can live in your cloud. Plug in your own by implementing the contract.
- microVM
Firecracker or Cloud Hypervisor, the default
- Containers
For trusted code
- WASM
For agent tools
- Browser and desktop
For agents that work with UIs
- gVisorplanned
Extra kernel isolation
- GPUplanned
For ML workloads
type Runtime interface {Create(…)Exec(…)Snapshot(…)Fork(…)Promote(…)Destroy(…)}
deploy · where it runs
Your servers. Your cloud. Your rules.
The core is open source under Apache-2.0. You can move from managed to self-hosted.
Self-hosted
The open core on your own servers, including air-gapped sites.
Apache-2.0
BYOC
Last.Backend runs the control plane; runtimes and data stay in your AWS, Google Cloud, Azure, Yandex Cloud or Selectel account.
your data stays with you
Managed
Last.Backend Cloud, with nothing to deploy.
move to self-hosted any time
- MicroVM runtimeearly access
- Container runtimeearly access
- MCP serverearly access
- Egress policyearly access
- Secrets brokerin progress
- Approvals (Slack, CLI)in progress
- gVisor runtimeplanned
- GPU runtimeplanned
- SOC 2planned
lastdevops × last-backend
How LastDevops fits in
Last.Backend was founded by Alexander Kaloshin, a co-founder of LastDevops. In 2014 it was his own container orchestrator; today it is a platform for AI agents. We roll it out the way we run all infrastructure: as code, monitored and on call.
- 01
Deploy it in your cloud
Self-hosted or BYOC, on your servers or in your cloud account, built with Terraform and handed over with docs.
- 02
Wire it into your workflow
Connect it to your Git hosting, CI/CD, monitoring and SIEM, so agents ship through the same release path as your team.
- 03
Write the policies
Egress, secret access, budgets and who approves production, kept in a repository and reviewed like code.
- 04
Pick the first agent workloads
We help choose what to hand to agents first and run a pilot with a metric agreed up front.
quick start
Add it to Claude Code
The Last.Backend MCP server is one command away.
$ claude mcp add lastbackend -- npx -y @lastbackend/mcpReady to give your agents a way to production?
Request early access at lastbackend.com, or tell us about your infrastructure and we'll help you roll Last.Backend out.