Skip to content

Overview

Inline Agents turns "run this workflow" instructions in an Obsidian note into a button. Click it, and it fires Claude Code or OpenCode at the note, in a live terminal right underneath — or, if you'd rather not watch it work, quietly in the background with the result dropped straight into the note when it's done.

Or hand the prompt to an agent you already run elsewhere. With target: <name>, the button sends its prompt over HTTP to an agent runtime — OpenClaw, Hermes Agent, or any OpenAI-compatible endpoint — and the work carries on there. See Agent Runtimes.

Built to replace plain-text "Run features/learning-artefacts/teacher-artefact-leveller.md" instructions scattered across a note system with something clickable.

Drop a fenced code block with the agent-button language tag anywhere in a note:

```agent-button
text: Draft this week's report
prompt: Draft a weekly report for {{file.basename}}, using the notes tagged #weekly-log from the last 7 days.
agent: ClaudeCode
model: myMainModel
showTerminal: true
completion: manual
```

That's it — every field except text and prompt is optional. Click the button and you get a real terminal, right in the note:

Buttons rendered in a note, idle

What you could use it for

  • A "Run sufficiency check" button under every artefact note, instead of remembering the exact command.
  • A background summarizer that appends its findings to the bottom of a daily note and never opens a terminal you have to watch.
  • Different buttons on the same note routed to different agents or different models — a fast local model for a quick pass, a bigger one for the real thing.
  • A "Hand to my agent" button that sends the note to OpenClaw or Hermes Agent, so the work runs with that agent's own memory and tools while you carry on in Obsidian.
  • Prompts that pull real context from the note — frontmatter status, tags, the vault name — instead of static text you have to edit by hand every time.

How it works

Each agent-button code block is its own self-contained widget, rendered via Obsidian's own registerMarkdownCodeBlockProcessor. Clicking the button:

  1. Renders the prompt: template against a snapshot of the note (file path, frontmatter, tags, the vault, today's date — see Context and Templating).
  2. Resolves which agent binary to run and, if a model: line is set, which model to pass it — see Agents and Models.
  3. Spawns the CLI behind a real PTY and feeds its output into a live xterm.js terminal in an accordion under the button — a real terminal, not a plain text log. Type into it exactly like a normal terminal whenever the agent isn't auto-approving.
  4. Once the process exits (or, in responseEnd completion mode, once the agent finishes its reply), the button returns to idle — and if agentOutput: file is set, the result is written straight into the note:

Agent output written into the note

A button with target: <name> skips steps 2–4: the rendered prompt goes to the named agent runtime over HTTP, the button shows "Sent to <name>", and you read the reply on the runtime.

Where to go next