AgentUse v0.19: a native Mac app and learnings that earn their place

AgentUse v0.19 is out, spanning 0.19.0 through 0.19.3. This release brings AgentUse to the Mac as a native app, replaces automatic learning capture with a vetted, opt-in pipeline, and moves provider setup and schedule control into the dashboard so a new user can go from an existing project to a running agent without editing files.

Get it as AgentUse for Mac, which bundles the runtime and dashboard, or run npx -y agentuse@latest setup in a terminal with Node.js 22+. Existing installs update through the Mac app's Settings → About, or by running the same command again.

AgentUse for Mac

AgentUse for Mac is a native menu-bar app with the complete runtime and operations dashboard bundled inside. Schedules, approvals, and active runs keep working when the window is hidden. The app can start quietly at login, attach to a local daemon you already run, or manage its own server. Native Settings cover server status, logs, login behaviour, notification categories, command-line access, and an optional global Dashboard shortcut.

The app ships its own CLI. Settings can add an ~/.local/bin/agentuse launcher without touching other installations, and lists every agentuse on your login-shell PATH in the order the shell resolves them, so a stray global install can no longer redirect first-agent setup to the wrong runtime profile. Quitting from the menu bar warns before it stops a server it manages. Closing the Dashboard just hides it.

Since 0.19.1 the app updates itself from Settings → About, and since 0.19.3 it downloads updates quietly and asks when they are ready. Restart Now installs through the normal clean shutdown path. Later keeps you working, and an ordinary quit never installs anything behind your back. App-owned servers are tied to the app's lifecycle: a crash triggers the server's graceful shutdown, and a survivor is reconciled on the next launch only after its recorded owner is confirmed dead. Servers you started in Terminal are left alone.

One onboarding journey on Mac, Web, and CLI

All three surfaces now guide a new user through the same first-agent flow. Setup creates and registers a project, runs a credential-free sample, confirms a runtime provider is ready, and prepares a structured handoff for Codex, Claude Code, Cursor, or another coding agent. Provider login stays in Terminal. OAuth URLs, authorization codes, and API keys never pass through a chat window.

0.19.2 extends this with project discovery. The dashboard attaches or creates a project, connects a provider, scans a sanitized read-only view of the project for evidence-backed recurring work, and creates the reviewed agent through a visible AgentUse session. The proposed schedule stays paused until you test and enable it. Discovery accounts for agents that already exist so it does not propose duplicate responsibilities, refuses to follow manifest symlinks outside the project, and redacts common secret assignments and token formats before anything reaches the model.

Model providers are configured from the dashboard too: Anthropic and OpenAI by OAuth or API key, OpenRouter and OpenCode Go keys, and custom OpenAI Completions, OpenAI Responses, or Anthropic Messages endpoints. Custom endpoints discover and save their models and verify the selected protocol with a real minimal completion.

Learnings are vetted, and human learning is deliberate

The old learning pipeline validated captured rules for shape, not substance. Nothing checked whether a rule duplicated the agent's own instructions, contradicted them, or was even supported by the run it came from, and whatever capture produced became active immediately. Worse, the evaluator saw only the first 3,000 characters of the agent contract, so on any real agent it routinely "rediscovered" explicit rules as new learnings.

Every candidate now passes three checks against the complete effective instructions:

• Duplicate: it restates a rule the contract or an active learning already carries.
• Contradiction: it conflicts with either, quoted in the verdict.
• Grounding: the session trace must actually contain what the claim describes.

Model-authored candidates that fail duplication or grounding are rejected. Contradictions are quarantined, not dropped: stored with the conflicting text named, visible in agentuse learnings <file>, the serve learnings panel, and agentuse doctor, and never injected. A human correction is never silently discarded.

Human learning is now explicit. A review comment becomes durable only when the reviewer checks Learn from this comment, uses --remember, or adds a learning by hand. Unchecked comments revise the current run only. Automatic capture is opt-in per channel:

learning:
  capture:
    custom: "Capture formatting corrections the reviewer makes to the brief"
    addons: [tool-errors]

capture.custom scopes the built-in evaluator with your text. capture.agent replaces it with an agent file whose output still passes the common vet. tool-errors is the first typed channel: a failed tool call, a corrected call, and a confirmed success, verified structurally in code.

Learnings also carry the contract they were captured against. When you rewrite an agent's instructions, learnings vetted against the old text are held out as stale rather than injected unexamined, and the next capture or tidy pass re-vets them. Existing learnings files load without migration and keep their state.

Three changes are breaking. A bare agentuse serve no longer adopts the current directory as a project; use agentuse serve -C . or persist the directory under serve.projects, and startup prints that command when it finds agents but no configured project. learning: true no longer auto-learns from a run; it parses, warns once with the mapping, and applies only deliberate learnings. learning.criteria is removed and is a hard parse error naming the rewrite.

Schedules, profiles, and the dashboard shell

Schedules can be paused per deployment without editing agent source, from the agent page, the schedules dashboard, or the new agentuse schedules pause and agentuse schedules resume commands. The declared cron expression stays visible. Pause and resume are serialized across processes so concurrent requests cannot overwrite each other.

AGENTUSE_CONFIG_DIR and AGENTUSE_DATA_DIR give explicit profile boundaries. The config root holds config.json, .env, managed projects, and user-global plugins and skills. The data root holds credentials, sessions, learnings, schedules, and server state. The file-level AGENTUSE_CONFIG and AGENTUSE_ENV overrides still work but are deprecated for removal no earlier than December 1, 2026.

The dashboard gained responsive, persistent navigation: a drawer on small screens, a collapsible and resizable sidebar on desktop with ⌘/Ctrl+B, and back/forward, section shortcuts, and ⌘R reload in the Mac app. The model catalog and compatibility routing were refreshed for the current GPT-5.6, Claude, OpenRouter, OpenCode Go, and custom-endpoint surfaces, with reasoning effort normalized per model and transport.

v0.19 is the release where AgentUse stops assuming you live in a terminal. The runtime is the same. It now sits in the menu bar, sets itself up from a project you already have, and only learns what it can prove.