self-hosted control plane · keyless by default · per-agent DB included

Deploy an internal agent in one command

Every internal agent — hosted, metered, and scoped — without anyone touching a Kubernetes cluster.

The problem

Engineers write agents. Then what?

Agents ship as ad-hoc scripts, notebooks, or—worse—unsandboxed processes. Nobody knows how many are running or what they access.

IT has no visibility

Spend is unmetered. Tool access is unconstrained. When things break, there is no audit trail. Kill switches do not exist.

The gap is a compliance nightmare

Teams build in secret. IT finds out when something goes wrong. By then, the agent has already run wild.

How it works

01

Write with defineAgent

Your team writes an agent using the @helm/agent contract.

02

Deploy with one command

Run `npx helm deploy`. The CLI bundles and sends it to your control plane.

03

Auto-provisioned execution

Your agent runs in an isolated child process. A per-agent Convex DB provisions automatically.

04

IT governs everything

Dashboard shows scope, token spend, and per-agent kill switches. No agent runs without governance.

Architecture
@helm/core → cost & policy engine
@helm/agent → defineAgent contract
@helm/runtime → execution & metering
@helm/cli → deploy & auth

IT sees and controls every agent

This is what separates Helm from "just run it on a VM." IT controls what every agent can do, how much it costs, and when it stops.

Tool & domain scoping

Each agent runs with an explicit allowlist of which APIs and domains it can reach.

Per-agent spend caps

Set a hard token budget. When it hits zero, the agent stops. No exceptions.

Hard kill switch

One click in the dashboard stops any agent mid-run.

Full audit trail

Every execution logged. Token spend tracked. Who changed what and when.

Agent: invoice-processor

Budget: $10 / mo
Spend to date$3.47
Remaining budget$6.53
StatusActive

Free and open source

Helm is MIT licensed and self-hosted. There is no paid tier, no seat count, and nothing to talk to sales about.

Run it yourself

Clone the monorepo and run the control plane on your own infrastructure.

MIT licensed, no usage limits
No account, no key, no telemetry
Codex CLI by default, or bring your own LLM
Per-agent Convex database provisioned on deploy

Governance included

Every control ships in the repo. Nothing is held back behind a tier.

Per-agent tool and domain scoping
Per-run and monthly token budgets
Budget gate enforced before execution
Run history and per-agent kill switch

Built to be read

The repo is the documentation: architecture notes, research, and tests.

Unit tests across core, agent, runtime, and CLI
Architecture and market research in docs/
Contribution guide and issue templates
Open to PRs and design discussion
$git clone https://github.com/nikolas-sapa/helm
View the source on GitHub

Deploy your first agent

Helm is open source and self-hosted. Clone the repo to run it, or leave your email to hear when a release lands.