Getting started

The first time you open Pilog it walks through a short onboarding: confirm the global hotkey, sign in to GitHub, link a local repository, configure the Pi draft agent, capture a note, and generate your first draft. You can skip onboarding and return to any step from Settings.

1. 01
   Confirm or rebind the global capture hotkey. Default is Ctrl⇧Space. The hotkey is registered against your OS, so it works whether or not Pilog has window focus.
2. 02
   Connect GitHub (next step).
3. 03
   Link one local repository (the folder on disk you intend to file issues against).
4. 04
   Configure the Pi provider and model (further down this page).
5. 05
   Capture a quick note with the hotkey to confirm the loop works end-to-end.

Pilog signs into GitHub using the Device Flow. When you click Sign in, Pilog opens GitHub in your browser, you authorize the app, and the access token is stored in OS-backed credential storage (Keychain on macOS, DPAPI on Windows, the keyring on Linux). No personal access token is required.

The OAuth scope is the minimum needed to read repository metadata and create issues on your behalf. Sign out from Settings → Account at any time; the token is wiped from keychain immediately.

Pi is the local agent harness that reads your selected notes alongside the active repository and produces structured issue drafts. Pilog bundles the Pi runtime; you choose the provider and model. A cloud API is the default path; for inference that never leaves your machine, use a local model via Ollama (recommended) or another OpenAI-compatible local server.

In Pilog, open Settings → Provider & Model and:

1. 01
   Pick Anthropic from the provider dropdown.
2. 02
   Pick claude-sonnet-4-6 from the model dropdown.
3. 03
   Paste your Anthropic API key. Get one at console.anthropic.com/settings/keys . Create a workspace if you don't have one, then Create Key. Keys start with sk-ant-….
4. 04
   Click Configure Pi. The key is stored in OS-backed safe storage, separate from Pilog's SQLite database.

Already use the pi CLI on this machine? Click Import existing Pi config on the setup panel and Pilog will pick up provider credentials from ~/.pi/agent/auth.json without you re-entering anything.

Press Ctrl⇧Space from anywhere. A quiet markdown scratchpad opens; type the thought as it occurs to you. No fields, no repo selector, no priority picker. Save with CtrlS or simply close the window. The note lands in your inbox.

Notes are plain markdown. Links, code fences, lists, and inline backticks all render in the draft generator's context. You can edit, retitle, or delete a note from the inbox at any time.

Open the inbox (Ctrl1), select one or more notes belonging to the same linked repo, and press GD, or click Generate Drafts. Pi reads the notes alongside repository metadata, groups related ideas, and writes structured issue drafts with titles, bodies, suggested labels, affected files, and a confidence signal.

Every draft is visibly anchored to the notes that produced it. The reasoning summary is short by design; the raw source notes are never hidden behind an expander by default. If the agent grouped two notes you wanted to keep separate, you can split a draft and regenerate just that subset.

Settings exposes the full Pi model registry bundled with Pilog. Browse pricing and context windows at pi.dev/models, then use the same provider and model id here. Default for drafts: Anthropic claude-sonnet-4-6.

• Direct APIs: Anthropic, OpenAI, Google (Gemini), DeepSeek, xAI, Mistral, Groq, and others.
• Gateways: OpenRouter, Vercel AI Gateway, Amazon Bedrock, Google Vertex, Azure OpenAI, Cloudflare AI Gateway.
• Coding-agent sign-in: GitHub Copilot, OpenAI Codex, OpenCode, Kimi Coding (OAuth where supported).
• Custom & local: Ollama, LM Studio, vLLM, or OpenAI-compatible endpoints via ~/.pi/agent/models.json. See Local models (Ollama) below. Results vary with smaller models.

Pilog is local-first for your journal: notes, drafts, and run history stay in SQLite on your machine. Generate Drafts is the step that can leave it, unless Pi talks to a model running locally. Ollama is the recommended path; LM Studio and other OpenAI-compatible servers use the same models.json pattern with a different baseUrl.

1. 01
   Install Ollama and confirm it is running (the menu-bar app on macOS, or ollama serve on Linux).
2. 02
   Pull a coding-oriented model. For example:

   $ollama pull qwen2.5-coder:7b

   Larger models draft more reliably; smaller ones are faster on modest hardware.
3. 03
   Create or edit ~/.pi/agent/models.json so Pi can reach Ollama (use the model tag you pulled):

   {
     "providers": {
       "ollama": {
         "baseUrl": "http://localhost:11434/v1",
         "api": "openai-completions",
         "apiKey": "ollama",
         "models": [{ "id": "qwen2.5-coder:7b" }]
       }
     }
   }

   Ollama ignores apiKey; Pi requires the field. Full options are in Pi's custom-models docs.
4. 04
   In Pilog, open Settings → Provider & Model. Choose ollama and your model. Paste ollama (or any placeholder) into the API key field and click Configure Pi. Pilog stores it in OS-backed safe storage; Ollama does not validate it.
5. 05
   Generate a test draft. Traffic stays on localhost. Publishing to GitHub still requires network access and remains a separate, explicit step.

Which key or sign-in you need depends on the provider you pick. The onboarding path below uses Anthropic; for any other provider, open pi.dev/models and use that vendor's console (or OAuth, for coding-agent providers).

• Anthropic (recommended): console.anthropic.com/settings/keys → Create Key. Add credit at Billing. Typical draft generation costs a few cents per run.
• OpenAI: platform.openai.com/api-keys → Create new secret key.
• Google (Gemini): aistudio.google.com/apikey.
• OpenRouter: openrouter.ai/keys. One key reaches many catalog models.

API keys are stored in OS-backed safe storage, never in Pilog's SQLite database and never written to logs.

Open Settings → Provider & Model, change the selection, and click Change…. The change applies to new draft generations; runs in the agent run history retain the model they were generated with so the audit trail stays honest.