Helm

What it is

An employee writes an AI agent, runs one command, and gets hosting plus an auto-provisioned per-agent database. IT scopes what each agent can touch, watches every execution, and tracks token spend per agent — with a hard kill switch.

Architecture

• @helm/core — cost metering, policy engine (tools/domains/budget), bundle hashing
• @helm/agent — defineAgent contract, domain-gated fetch
• @helm/runtime — Codex runner, pluggable provider layer (Codex/Anthropic/OpenAI), isolated bundle execution with parent-proxied metering, Convex provisioner
• @helm/cli — init / deploy / auth / whoami
• Dashboard — Convex schema + Hono ingress

Current status

MVP built end-to-end and verified live against real Codex + Convex. 42 commits, 58 tests, security review on every commit. Marketing landing page live at helm-internal-agent-platform.vercel.app. Untrusted multi-tenant execution is gated behind a container backend (not yet built).

Stack

TypeScript monorepo (npm workspaces), Convex, Hono, Vercel Sandbox, keyless Codex CLI default with pluggable LLM backend.