v0.2 · agent environment control plane

Your agent stack needs one source of truth.

Initialize one Git-backed environment. Inspect what already controls your agents. Plan, install, update, and uninstall the agentic CLIs every high-agency builder relies on.

macOS + Linux Node.js 20+ inspect installer
curl -fsSL https://agentctl.caliberworks.co/install.sh | bash

Installs to ~/.local without sudo. Verifies the release checksum.

Representative local output plan first

$ agentctl agents status

agentctl agents status · version probes only

3/7 installed · 1 drifted

TOOLSTATECHANNEL
Codexinstalledbrew Claude Codeinstallednative OpenCodeinstallednpm Gemini CLImissingnpm Copilot CLIabsentnpm OpenClawabsentnpm Hermes Agentabsentnative

DRIFT Gemini CLI is desired but not installed

NEXT agentctl agents install gemini --dry-run

Canonical root: ~/.agentctl · Use --json for automation.

Codex Claude Code Gemini CLI OpenCode Copilot CLI OpenClaw Hermes Agent Cursor inspection Future adapters

Canonical by design

One root. Four control surfaces.

Start at ~/.agentctl, commit the human-readable intent, and keep machine observations explicit and disposable.

01

Canonical source

Initialize an idempotent Git repository with shared guidance, targets, and desired tool state.

agentctl init
02

Environment evidence

Inventory roots, resources, scripts, hooks, MCP surfaces, duplication, divergence, and risk.

agentctl inspect
03

Tool truth

Compare version-controlled intent with executable, version, channel, and verified machine receipts.

agents status
04

Lifecycle control

Preview exact static recipes, confirm explicitly, verify the result, and preserve configuration.

--dry-run → --yes

Agent Environment as Code

File sync is not semantic control.

A file can arrive at the right path and still be ignored, weakened, or interpreted under a different scope. agentctl keeps the evidence layers separate.

  1. 01 Desired canonical intent
  2. 02 Rendered target-native output
  3. 03 Filesystem observed in v0.2
  4. 04 Runtime target discovery
  5. 05 Behavior bounded evidence

v0.2 versions desired tool truth and observes executable truth. It never claims that static presence proves runtime discovery or model behavior.

The honest release

Useful now. Explicit about next.

v0.2.0 alpha

Shipping

  • Read-only multi-agent inventory
  • Duplicate and divergence analysis
  • Secret-key and executable-surface signals
  • Human and stable JSON output
  • Safe canonical repository scaffolding
  • Seven-tool catalog including OpenClaw and Hermes
  • Plan, confirmation, channel inference, and drift evidence
  • Configuration-preserving uninstall
  • Verified agentctl self-update and safe self-uninstall
  • Checksum-verified, no-sudo installation

Roadmap

  • Target-native semantic adapters
  • Capability-loss accounting
  • Content-addressed configuration plan and apply
  • Atomic rollback and drift records
  • Secret-manager references
  • Runtime and behavioral conformance

Trust is product behavior

Control should not require surrender.

NP

No surprise writes

Lifecycle commands show the exact plan and require --yes.

NS

No shell interpolation

Reviewed static arguments execute directly; arbitrary command strings are not accepted.

NC

No configuration purge

Uninstall removes the selected CLI distribution and preserves agent credentials and data.

NT

No telemetry

The CLI has no analytics endpoint, account requirement, or environment upload.

Read the threat model

Open source · Apache 2.0

Own the environment.
Then scale your agency.