Quickstart

1. Download the macOS build

The download page only shows a real button for a platform with an available package: macOS 14+ on Apple Silicon. It reads latest.json and prefers download_url or dmg_url when present; otherwise it falls back to a static link.

If macOS blocks the first launch as an unidentified developer, right-click Mycel in Finder and choose Open. That is not a notarization promise — the release pipeline still needs the owner-side Apple and release secrets run end to end.

2. Open your Markdown or Obsidian vault

1. On first launch you can use the starter vault, or open your own knowledge-base folder from the home entry points.
2. Mycel reads the selected local Markdown folder. Dot directories such as .obsidian/ are not treated as content notes.
3. The starter vault’s AGENTS.md tells external agents that Markdown is the source of truth, and that content edits should go through proposals — never silent direct writes.

3. Learn Today first

1. After opening a vault, land on Today. Read the greeting, the 30-day growth curve, and the Waiting on you cards so you know where Mycel surfaces decisions.
2. The current navigation is grouped into five buckets: Today, To handle, Knowledge, Tasks, and Agent.

4. Connect your first agent

The desktop package ships a native MCP stdio entrypoint: chunfeng mcp. Most MCP clients share the same JSON shape; Codex uses TOML.

chunfeng is Mycel’s internal executable name (crate / app binary). Copy it exactly in MCP configs — do not rename it to mycel or the connection will fail.

{
  "mcpServers": {
    "chunfeng": {
      "type": "stdio",
      "command": "/Applications/Mycel.app/Contents/MacOS/chunfeng",
      "args": ["mcp"]
    }
  }
}

[mcp_servers.chunfeng]
command = "/Applications/Mycel.app/Contents/MacOS/chunfeng"
args = ["mcp"]

The in-app MCP settings panel shows client registration and bridge status. Still verify the written MCP config on your own machine once.

5. Pick an agent and model

1. Open the Agent workspace inside Mycel and choose one installed local agent.
2. If that agent supports local model discovery, choose a discovered model. If not, Mycel falls back to a static suggestion list.

6. Submit the first proposal

1. Give the agent a small task — for example, “add a short summary to the current note and submit it with propose_edit; do not edit the file directly.”
2. Open To handle → AI Proposals, or use the Today action card CTA, then expand the pending proposal and review the diff, target path, agent, reason, and risk markers.
3. Approve when it looks right; otherwise reject with a reason. Rejection goes through the backend reject_proposal command — the frontend does not silently rewrite Markdown status.
4. To revert an approval, use Undo in the Proposal Queue. Undo depends on the previous-content snapshot recorded at accept time and on git history.

7. Next steps

After the approval loop works, read Capture and Governance. If you only need to self-host the relay, go straight to Self-host relay.