$ lumen --help

Help

Scannable cheat-sheet. Every control in the app has a tooltip — hover for the full detail. This page is the index.

Quickstart

1. Sign up at /signup (14-day trial, no card).
2. Add an API key at /settings — or skip and use Claude Code subscription mode.
3. Click + New workspace on /library, or + Import from GitHub to clone a repo.
4. Type your intent in the left pane → Send.
5. Click Run latest → to dispatch the result through Claude Code on your machine.
6. Review the diff on the subgoal card. Nothing auto-commits.

Glossary

Two panes: Plan and Run

• Left (Plan) — chat with Lumen. Design / refine / review. No code runs here.
• Right (Run) — execute. Subgoal cards stream live status, tool calls, output, and diffs.
• Run latest → (composer) — sends the most recent assistant message to the right pane.
• Send to Run → (per-prompt button) — same, but for a specific message.
• Dispatch — spawns Claude Code with the prompt + workspace codebase. Subscription mode by default. Files land on disk; no auto-commit, no push.
• ← Load latest result — pulls the most recent subgoal output back into the chat composer for refinement.
• Cancel — SIGTERMs the running subprocess; subgoal flips to failed with Retry / Skip / Discuss.

Hover any of those buttons in the app. The tooltip carries the same info this list does — you don't need to come back here.

Blueprints

A blueprint is a structured spec doc. 14 sections, in render order:

• Overview, Goals, Non-Goals, Personas, User Stories, Architecture, Data Model, API Surface, Capabilities, Security Model, Performance, Milestones, Open Questions, Decisions.

Create

In chat: "start a blueprint for X" or "save this as a blueprint."

Freeze

In chat: "freeze the blueprint as v1" (optional summary). Snapshots the draft as an immutable version (v1, v2, …). Auto-flips workspace phase to blueprinting if it hasn't graduated. Frozen version shows up on the Library card with a v# badge. Draft stays editable.

Browse

Click the N blueprint(s) ▾ chip on any Library card to expand the list of blueprints + versions.

Library and phases

Every workspace lives in /library. Filter by phase, search by name / path / blueprint title. Per-card three-dot menu: open / open-in-editor / change phase / archive / delete.

• Blueprinting — spec'ing. Auto-set when you freeze a blueprint without dispatching.
• In progress — building. Auto-set on first dispatch; default for new workspaces.
• Shipped — deployed at least once.
• Paused — on hold (customer-set, never auto).
• Archived — hidden from default lists. Data stays; unarchive any time.

Providers

• Anthropic (Claude) — console.anthropic.com. Default. Sonnet 4.6 routine, Opus 4.7 deep, Haiku 4.5 short factual (smart router picks if no model pinned).
• OpenAI — platform.openai.com. GPT-5 general, o3 reasoning-heavy.
• Google — aistudio.google.com. Gemini 2.5 Pro (long-context), Flash (cheapest).
• xAI — console.x.ai. Grok 4 (real-time web data).
• Ollama — ollama.com/download. Local, free, no rate limits. Default auto-fallback when cloud providers 429.

Consumer subscriptions are not API access. Only Anthropic lets you use Claude.ai Pro/Max through Claude Code subscription mode (no separate key). ChatGPT Plus and Gemini Advanced do not include API access — you need a key.

MCP (outside services)

Each MCP server adds its tools to every chat turn. Three presets are one-click:

• GitHub — create PAT → save under service name github in /settings → chat Settings → + GitHub preset → Test connections.
• Cloudflare — terminal: npx mcp-remote https://bindings.mcp.cloudflare.com/sse → browser OAuth → preset → Test.
• Vercel — same flow with npx mcp-remote https://mcp.vercel.com/sse.

Custom MCP server: chat Settings → + Add custom MCP server. Env values starting with vault: resolve to a key saved at /settings.

Tools

Read context

repo_search, web_search (Brave), web_fetch, system_scan (installed software).

Workspaces & blueprints

list_workspaces, list_blueprints, list_blueprint_versions, list_plans, get_blueprint, get_plan, create_blueprint, patch_blueprint, freeze_blueprint.

Self-improvement (codebase workspaces)

git_commit_and_push (never force-pushes), latest_commit_sha, get_ci_status, get_ci_failures.

Closes the loop: write a fix → ship → watch CI → react to failures — no bouncing through you.

Autonomous loop

lumen autonomous --workspace <slug> --intent "<what to improve>" --max-iter 3

Codebase workspaces only. Never force-pushes. Stops on dirty tree, max-iter, or all-CI-green.

CLI

lumen project create <name> [--from-github <url>]
lumen project use <slug>
lumen project ls
lumen chat
lumen plan <intent>
lumen dispatch <plan-id>

--from-github accepts full URL, owner/repo, or SSH. lumen --help for everything.

Cost

• Lumen software: $20/mo or $149/yr.
• AI inference: BYOK — provider bills you directly, no Lumen markup. Or use Claude.ai Pro/Max via Claude Code subscription mode (no API key).
• Trial: 14 days, no card.

FAQ

Forgot password?

/forgot.

Daemon needs reload?

Auto on new build (file watcher). Manual: chat Settings → Reload daemon. Last resort: Restart-Service LumenDaemon.

Rate-limited?

SDK stream fallback is off by default (fallbackProvider: none). Opt in to Ollama or another provider in workspace settings if you want a mid-stream downgrade when an API call fails.

Privacy?

Provider sees what you send per their policy (Anthropic / OpenAI / Google don't train on API traffic by default). Lumen persists prompts + responses to your tenant DB. Sign out invalidates session. Delete account from /account wipes data.