Skip to content

Latest commit

 

History

History
104 lines (86 loc) · 5.54 KB

File metadata and controls

104 lines (86 loc) · 5.54 KB

1Helm product vision

1Helm gives an AI employee a place to work, not merely a chat box in which to describe work. Every ordinary channel owns exactly one resident identity, one durable workspace, one memory namespace, and one persistent private Linux computer: an Apple container machine, unprivileged LXC, or private WSL 2 distribution according to the host platform. The human is the Captain. Skipper is the workspace-wide operator for host, fleet, credentials, connections, and cross-channel boundaries.

The outcome contract

A resident owns the user's requested outcome. Routine inspection, research, downloads, package installation, configuration, commands, retries, file work, and reversible implementation choices happen autonomously inside its computer. It does not turn executable work into a tutorial for the user.

When work crosses the channel boundary, the resident calls Skipper directly. Skipper performs the boundary operation and returns the same thread to the resident, which verifies the original end-to-end outcome. The Captain is involved only for consequential human judgment, credentials that only they can provide, external authority, or an irreversible commitment.

The runtime gives the model a compact factual capability map: its identity, channel, Linux computer and /workspace, callable tools, memory, skill arsenal, and authority scope. It does not inject a behavioral manifesto or grade and rewrite the model's prose. Tool and server implementations enforce isolation, ownership, credentials, destructive-action confirmation, and human-only boundaries.

If work depends on time passing, it becomes a persisted obligation. 1Helm does not equate a model ending a turn with background execution. Follow-ups and recurring workflows have due times, retry state, computer wake obligations, and observable completion or failure.

Durable growth

Residents start with a substantive operational arsenal, not a handful of generic prompt snippets. The model sees concise skill metadata and chooses when to load one full procedure. Skills define activation, execution, authority boundaries, recovery, retained state, and verification. Trusted external procedures may be discovered and installed automatically only through bounded, provenance-pinned, scanned, hashed ingestion beneath runtime policy. Community popularity is not trust.

When a resident completes a repeatable workflow that the arsenal does not cover, it can crystallize the proven procedure with concrete evidence. Generic advice, unverified guesses, secrets, incident-specific personal data, and instructions that weaken the runtime are rejected. Corrections, curated memory, skills, artifacts, and obligations compound without binding the identity to one model provider.

Raw prior-session transcripts remain an on-demand resident capability rather than prompt baggage. Each resident can search its own channel archive by meaning, exact text/date, or recency and hydrate a selected session in full. The message database remains authoritative; the resident's separate Mnemosyne database is a scoped semantic index. Models receive the tools and choose when they are relevant without a behavioral decision tree.

Computer and connection boundaries

On supported Macs, each resident computer is a real Apple container machine with the Mac home directory unmounted. Linux systemd hosts use an unprivileged LXC with subordinate host IDs and private networking. The accepted Windows implementation uses one private WSL 2 distribution with Windows-drive mounts and interop disabled. The resident's command tools and the channel Terminal share the same /workspace. Skipper manages CPU, memory, disk, wake/sleep, repair, update, archive, restore, and deletion safety.

Provider accounts and native connections remain host-owned brokers. Residents receive narrow operations, never raw OAuth tokens, Photon project secrets, or a private host database. Gmail currently supports scoped account listing, search, message retrieval, and draft creation. Photon/iMessage currently treats text as the reliable contract; richer attachment fidelity remains under verification.

Product truth

  • https://1helm.com is the standalone product and documentation site.
  • https://demo.1helm.com is a separate public sandbox, not the product site and not a dependency of an installed workspace.
  • Apple Silicon macOS and Linux systemd are public host products with one isolated Linux computer per ordinary resident.
  • The native Windows x64 + WSL implementation and real-host lifecycle are accepted, but its public installer is withheld until an Authenticode signing identity is available.
  • 1Helm is self-hosted and open source. A hosted control plane, mobile clients, blind community-skill execution, a signed public Windows installer, and a native Linux desktop shell are not shipped.

What 1Helm borrows—and what it does not

Hermes demonstrates the value of a capable built-in skill library and agents that act before escalating. Buzz demonstrates the value of honest capability matrices, screenshot-led explanation, agent-first JSON interfaces, first-class workflows, bounded stop hooks, outcome-first activity, auditable operations, and public benchmarks. 1Helm adopts those durability patterns where they strengthen outcome ownership.

1Helm does not adopt a protocol, event model, media feature, or git substrate merely for parity. Its core bet is the compounding resident world: one identity, one private computer, one workspace, and years of accumulated memory, skills, corrections, artifacts, and obligations—with Skipper automatically handling every boundary.