# Muse Spark 1.

> Source: <https://promptcube3.com/en/threads/8631/>
> Published: 2026-09-02 23:08:46+00:00

# Muse Spark 1.

If you've been living under a rock like I was until last week, Muse Spark is a self-hosted AI agent toolkit that's been quietly eating Rust/Cargo attention. Version 1.3 landed on Sept 2026 and the changelog is worth skimming, especially if you're tired of

Still, if you're running a local LLM stack and want something between raw llama.cpp and full Copilot, Spark 1.3 is the most pragmatic middle ground I've found. The HN thread has 197 comments (link above) — skim the first 30 and you'll catch the recurring complaints and workarounds faster than I did.

[Claude Code](/en/tags/claude%20code/)'s cloud dependency or Copilot's context limits.The headline for me: native streaming support for tool calls. Finally, no more waiting for the entire JSON blob to dump before your terminal updates. Here's the snippet I added to `.musespark/config.yaml`

to flip it on:

```
agent:
  streaming: true
  max_tokens: 4096
  model: "llama3.1:8b"
```

[Cursor](/en/tags/cursor/) devs will recognize the pattern — it's the same incremental rendering trick that makes their inline edits feel less janky.

The bigger win is the new plugin API. Spark 1.3 lets you register external tools without recompiling the core binary. I hooked up a custom SQL formatter and a dead-link checker with barely any boilerplate. Example plugin manifest:

```
{
  "name": "sqlfmt",
  "entrypoint": "./sqlfmt_plugin",
  "capabilities": ["format", "lint"]
}
```

Drop that in `~/.musespark/plugins/`

, restart the agent, and it just shows up in `--list-tools`

.

A few gotchas coming from [Claude](/en/tags/claude/) Code:

- No built-in git diff summarization. I patched it with a tiny shell plugin that runs
`git diff --stat`

before each commit suggestion. - The default prompt templates are sparse. I ported over a couple of my favorite Cursor rules by hand-editing
`~/.musespark/prompts/base.md`

. - Memory management still feels rough around the edges. I had to cap context length manually or it'd happily eat 2GB on a single refactor task.

Still, if you're running a local LLM stack and want something between raw llama.cpp and full Copilot, Spark 1.3 is the most pragmatic middle ground I've found. The HN thread has 197 comments (link above) — skim the first 30 and you'll catch the recurring complaints and workarounds faster than I did.

[Next I haven't searched for a stack trace on Google in years →](/en/threads/8591/)

Free AI toolbox — all free to use

Step-by-step guides and pitfalls for this path are in

[an AI side-hustle playbook](https://tanyan888.com/), with plenty of directly applicable cases.## All Replies （4）

T

Finally tried it last night. The setup was way smoother than I expected for self-hosting.

0

S

R

Does it handle custom WASM modules for the agent logic, or is it strictly Rust?

0

Q

Don't forget it also has built-in support for vector DB integration, which helps a lot.

0