Your AI agents, inside your vault.
Run Claude Code, Codex CLI, opencode or Antigravity CLI in a real terminal next to your notes. Start them with one click, give them one set of instructions, and let them drive Obsidian.
~/Notes $ claude --continue
Resuming session in ~/Notes…
> open today's daily note and add the meeting summary
● obsidian daily vault="Notes"
One-click launch
Start an agent, continue the last session or skip permission prompts from a button, the command palette or a hotkey. Commands are only typed while the shell is idle.
One AI core
Write your vault rules once. They are linked to CLAUDE.md, AGENTS.md and GEMINI.md, and shared skills appear in every agent.
Agents drive Obsidian
Through the official Obsidian CLI, agents can open today's daily note, switch view modes or run any command.
Presets for your agents
Every preset is plain text you can edit. Claude Code and Antigravity CLI are in the menu out of the box; add Codex CLI and opencode with one button in the settings.
| Claude Code | Codex CLI | opencode | Antigravity CLI | |
|---|---|---|---|---|
| New session | claude | codex | opencode | agy |
| Continue | claude --continue | codex resume --last | opencode --continue | agy --continue |
| Skip prompts ⛨ | --dangerously-skip-permissions | --dangerously-bypass-approvals-and-sandbox | --auto | --dangerously-skip-permissions |
| Instructions | CLAUDE.md | AGENTS.md | AGENTS.md | GEMINI.md, AGENTS.md |
| Skills | .claude/skills | .agents/skills | .agents/skills, .claude/skills | .agents/skills |
⛨ Use with care: commands that skip permission prompts let the agent run commands and edit files without asking. They are marked in the menu and in the settings.
Local and cloud models with Ollama
Add the Ollama presets to start Claude Code, Codex CLI or opencode with a local or Ollama cloud model. The agent keeps its tools: it still edits notes, runs commands and uses the shared skills.
- Choose a model that supports tools and set a context length of 64k or more. Plain
ollama runonly chats. - You pick the model the first time; later launches reuse it.
- Models ending in
:cloudrun on ollama.com.
| New session | Continue | |
|---|---|---|
| Claude Code | ollama launch claude | ollama launch claude -- --continue |
| Codex CLI | ollama launch codex | ollama launch codex -- resume --last |
| opencode | ollama launch opencode | ollama launch opencode -- --continue |
Let agents control Obsidian
Turn on Settings → General → Command line interface and run Bootstrap AI core. The included obsidian-cli skill teaches every agent how to use it, and to ask before anything destructive.
| Ask the agent to… | It runs |
|---|---|
| Open today's daily note | obsidian daily |
| Toggle reading view | obsidian command id=markdown:toggle-preview |
| Toggle source mode | obsidian command id=editor:toggle-source |
| Run any command | obsidian command id=<id> |
Install
- In Obsidian, open Settings → Community plugins → Browse and search for “Agent Terminal” (once listed). Or download
main.js,manifest.jsonandstyles.cssfrom the latest release into.obsidian/plugins/agent-terminal/. - Enable the plugin and open a terminal from the ribbon or the command palette.
- Allow the one-time download of the terminal engine, then press 🤖 and pick an agent.
Good to know
- Desktop only: Windows, macOS and Linux (x64, arm64), Obsidian 1.13.1 or later.
- Install the agents themselves as usual; Agent Terminal starts them.
- The only network request is the terminal engine (node-pty, 50–800 KB), downloaded once after you agree and verified with SHA-256. No data is collected.
- Before your own
CLAUDE.mdor skills, or an editedai-core.md, are replaced, a dialog lists them and asks first. Each one is kept as.bak, never deleted, and nothing changes if you cancel. - Open source under the MIT license.