Integration with an AI assistant

I just wanted to share how I've been using some of the new SilverBullet features. For the last month or two I've been using Hermes as an AI assistant and was interacting with it through my personal laptop.

When I saw the new Authorship functionality it occurred to me I could use SilverBullet as the communication medium, particularly as I quite often have Hermes comment on things within my space, and across the different devices I access SB with.

How it works:

  1. I can mention @hermes anywhere in my space and it will poll every couple of minutes until it sees itself mentioned, at which point it takes the context of the page and begins work, emitting an in-flight message.
    Screenshot 2026-09-21 at 14.16.42

  2. Once it's finished working, it adds its reply as a bullet. This mimics a threaded conversation somewhat.

  3. It will mention me and a listener sends a flash notification so I don't need to keep the Mention Inbox open.

  4. If I want to, I can continue the conversation in a nested list so the conversation is routed back to the original worker (maintaining the context).

  5. If Hermes ever needs my approval or an answer, it will raise this in a Comment Block. I can click Resolve to indicate my approval or to accept the default/recommended answer; otherwise I select or write my own and sign it.

  6. Hermes uses Admonitions to notify me of problems (with scheduled tasks, for example) on my daily journal page (where I spend most of my time), as well as sending a flash notification.

  7. Because of Revisions and Accounts, Hermes's edits are all correctly attributed and recorded.

Thanks to @zef for probably foreseeing a lot of similar use-cases and making the HTTP and Runtime APIs very good!

I'm hoping to package this up for the community once I've used it for longer.

I've been playing with Hermes too and have been thinking about this exact use case :wink: I'm currently cooking up some "skills" that should make this even easier (interacting with SB spaces, but also pulling in comments addressed at e.g. an agent via the sb CLI.

Hello, thanks for sharing!

Quick question, does your agent modify the plain text files directly over the command line? If so, are you using the "Managed" git version control option provided by SilverBullet 2.11?

It's using the HTTP API and mostly conditional writes. Yep, I'm using the managed mode.

This seems absolutely brilliant. I wish I was as smart as you guys so I could do something like this myself!

Very much a work in progress, but perhaps already helpful in this form: GitHub - silverbulletmd/skills: Agent skills for SilverBullet · GitHub (the mentions related skill do depend on a new API I just introduced on the edge builds)

Not sure if you use Hermes locally or on some VM, but installing SilverBullet+ alongside it with the sb installed alongside allows for some cool workflows. You can indeed also make it work with the remote API on a SilverBullet server.

AI gets the credit, I merely directed it (and questioned some of its decisions...) :smiley:

If Hermes is mcp-friendly, you might like this:

It uses the same runtime api that sb uses, but also exposes custom lua tools from the space. Admittedly the main reason I'm enjoying it is that I can have agents execute lua queries directly now.

The direction I’m thinking is that the CLI eval sub command should allow you/it do whatever it needs. Do you still need an MCP?

tbh I think the original reason I needed mcp was because I wanted to use claude desktop which needed mcp instead of a cli.

In a janky way, it also let me register mcp servers in SB and then share the same mcp tools from SB to other agents/clients instead of configuring each one. There are definitely tools that do that better though :smiley:

I prefer to use CLI in most cases especially now that I use remote control I don't need an MCP. Similar to @justyns the only reason I used it is for Claude Desktop.

Claude Desktop can run CLI tools also, no? At least in Claude Code mode? Or are you referring to the general chat section?

Related to this topic, my "web observer" just found this:

That's great to hear! I think the newest release is the deal breaker and I've updated my OpenClaw image right away. I use SilverBullet along OpenClaw to offload summaries, market research and monitoring and it is the single UI for me to read its output. This way SilverBullet becomes a channel.
FYI. I combine it with CloudFlare to restrict direct access to it.
Thanks for inspiration.

An argument for MCP over CLI:
It is a purpose built format specifically to make it easy for AI to extend their capabilities. LLMs have been trained on it so it is reliably understood. Capabilities are discoverable and have documentation attached. The permission model is better; clients typically offer clear ways for users to enable/disable specific tools or the entire server, and you can pre-allow specific tools, without needing to worry about leaking permissions by using overly permissive regex to match your bash command. You don't need to install the CLI in your local env, you can just connect the remote MCP server.

I may be in the minority here, but I still think that MCP holds value. I should give those plugs a look!