Configuration
Elowen keeps startup configuration separate from runtime settings. Environment variables establish the local process boundary; the Settings workspace persists operational choices and saves normal changes as you make them. Secrets are never returned to the browser after storage.
Startup environment
Set environment variables before starting the relevant process; the defaults are intentionally local-first. The full variable table — daemon port and bind address, database path, bootstrap credentials, logging, and the CLI client variables — lives in Production & Updates.
elowen setup is the preferred way to configure a local installation. Use elowen doctor to check the resulting daemon, provider, memory, and task readiness.
Settings sections
The Settings surface follows one stable order. It is administrator-controlled; user-level preferences and access live under Account and Users.
| Section | What it controls |
|---|---|
| System | Version/readiness, service controls, automatic updates, and login-token lifetime. |
| Elowen AI | Agent name, provider accounts, model context windows, max steps, runtime limits, and stale-conversation auto-cleanup. |
| Models | Enabled executor presets, custom entries, and model notes for planning. |
| CLI Agents | External coding CLI binary, arguments, permission behavior, and resume behavior. |
| Data | Explicit administrative maintenance and destructive cleanup. |
| GitHub | Write-only credentials and PR-workflow defaults. |
| Autopilot | Mission defaults, planning/review roles, TDD mode, and PR defaults. |
| Plugins | Installed capabilities, marketplace, and plugin-owned configuration. |
| Memory | Embedding and categorization models. |
Elowen AI and limits
Elowen AI providers can be OpenAI-compatible or Anthropic API-key entries, or supported OAuth accounts. The Web UI shows provider/model metadata and whether a secret exists; it never reads a stored API key back into the client.
The same section contains guarded runtime limits for tool-output display, elicitation waits, memory recall, autonomous goals, and live channel sessions. Values are validated and clamped by the daemon, so a stale or malformed client cannot persist an unsafe bound.
OAuth subscription usage
Each connected OAuth provider — Anthropic, GitHub Copilot, OpenAI Codex, and Kimi — gets a slim usage rail next to its account entry. The rail shows the fill of every usage window the provider exposes (typically the 5-hour and weekly windows) and shifts color as pressure builds: normal, warning at 70%, and danger at 90%. Hover a rail for a tooltip with the window's reset time.
Watch these rails before launching a long autonomous run. A provider at warning pressure can rate-limit the brain mid-task; pick a different model or wait for the reset.
Brain limits
The limits modal in Elowen AI exposes the eight runtime knobs that bound a single brain turn. Every value is clamped to a safe range, both in the form and again by the daemon.
| Limit | What it bounds |
|---|---|
| Tool output max lines | How many lines of a tool result the brain sees before truncation. |
| Tool output max chars | Character cap on the same tool output, applied alongside the line cap. |
| Elicitation timeout | How long the brain waits for an answer to a question it asks you. |
| Memory recall count | How many memories are recalled into a turn's context. |
| Memory recall chars | Character budget for the recalled memory block. |
| Goal turn budget | Total turns an autonomous goal may spend before stopping. |
| Goal max turns | Hard ceiling on turns for a single goal step. |
| Channel session cap | How many live agent sessions one channel may hold at once. |
Models and CLI agents
Models is the workspace catalog for task executors. It controls which known presets are visible, custom model entries, and notes that help planning select a suitable executor. Per-user allowed executors only narrow this workspace-wide ceiling.
CLI Agents configures the external programs Elowen can spawn. For each configured program, adjust the executable, arguments, permission behavior, and whether a restarted task should resume the prior external session. The embedded Elowen brain is configured in Elowen AI, not as an external binary.
Autopilot and GitHub
Autopilot holds mission-wide defaults: worker executor, autonomy, maximum concurrent sessions, planner/overseer choices, optional completion review, TDD mode, and PR behavior. A new mission can override its pilot and overseer without changing the global setting.
GitHub settings keep credentials write-only and define the PR workflow defaults: enabled state, base branch, auto-open behavior, and an optional verification command. Project and mission settings can override the PR enabled state. See Projects & Workflow.
Plugins and Memory
Plugins render their own config from elowen-plugin.json; use their documentation and help affordances for specialized fields. The Memory section chooses the embedding and categorization models. API-key and OpenAI-compatible providers can supply those credentials without duplication; OAuth-only accounts cannot, because they do not expose an embedding endpoint.
Leaving embeddings unconfigured keeps memory usable with keyword retrieval. Reindexing or recategorizing performs a deliberate background operation; it is not a hidden side effect of changing the visual UI.
Plugin editors: subagents and skills
Under Plugins, the subagent and skills plugins each ship a full editor so you can create, edit, and delete entries without touching the filesystem.
The subagents editor manages custom sub-agents. Each entry is a Markdown file: frontmatter declares the name, description, and tools mode — read-only, all, inherit, or an explicit custom tool list — and the body is the prompt the sub-agent runs with. The built-in explore and plan agents are read-only; you can inspect them but not modify them. Changes hot-reload, so a saved sub-agent is available to the next delegation immediately.
The skills editor follows the same pattern for skills: a Markdown body with frontmatter, listed and editable in place. See Skills for how skills are discovered and applied.
LSP diagnostics
Elowen can run language-server diagnostics after file edits, surfacing live type and lint errors in the session. It is on by default. The /lsp command in the CLI toggles it, and the choice persists across restarts. See Slash Commands for the command syntax.
Accounts and permissions
Account preferences are personal. Users and their project/model/tool access are managed separately. A setting being visible does not override the daemon's authorization checks; permissions and destructive actions retain confirmation boundaries.