Hephaestus

Hephaestus, stripped down to run as an Obsidian plugin: a chat pane inside your vault, talking directly to a local model server. No backend process, no Electron shell of its own — the plugin runs entirely inside Obsidian.

Features

• Chat with local models, with live token streaming and Obsidian's own Markdown rendering (code blocks, tables, callouts for free).
• Two backends. Ollama (/api/chat) and any OpenAI-compatible server — LM Studio, llama.cpp, vLLM, LocalAI, Jan — selected in settings. Both streaming and non-streaming paths are supported.
• Attachments. Images (to vision models) and text files, from your computer or picked out of the vault.
• Web search. Pluggable backend: DuckDuckGo (no setup), a self-hosted SearXNG instance, or the Brave Search API. Top pages are read and sources cited inline.
• Attach a web page. Paste a URL and its text is pulled in as context — no search engine involved.
• Thinking mode on models that support it (Ollama only).
• Note integration. Optionally send the open note as context, insert a reply at the cursor, or let the model append to the note itself — behind a confirmation prompt.
• Editable transcript. Edit any message, regenerate the last reply, or delete a message and its answer.
• Context gauge. A ring in the composer showing how full the context window is: green under 50%, yellow under 75%, red above.

What survives from Hephaestus, and what doesn't

Dev setup

npm install
npm run build      # type-checks and bundles to main.js
npm test           # runs the unit tests
npm run dev        # rebuild on save

Then copy (or symlink) this folder into your vault at <vault>/.obsidian/plugins/hephaestus/ and enable it in Settings → Community plugins.

Choosing a server

Settings → Hephaestus → Server:

• Ollama — http://localhost:11434, the default.
• LM Studio — http://localhost:1234. Start the server from LM Studio's Developer tab. Enter the base URL without /v1. The Test button checks reachability and lists the models it finds. Switching servers restores that server's default URL and clears the remembered model, since names differ between them.

A third option — a hosted "Cloud API key" provider — is built but not exposed; see CLAUDE.md.

Remote servers

Neither Ollama nor LM Studio has authentication, so never expose one directly to the internet. Use Tailscale/WireGuard (simplest — nothing in the plugin changes) or an SSH tunnel. A reverse proxy with bearer-token auth will not work yet: the plugin has no setting for an Authorization header.

CORS and streaming

Obsidian pages run from the app://obsidian.md origin, and model servers only accept browser requests from origins they trust. For streaming to work:

OLLAMA_ORIGINS=app://obsidian.md ollama serve

For LM Studio, enable CORS in the Developer tab. Without this the plugin falls back to Obsidian's requestUrl, which bypasses CORS but cannot stream — you get the full reply at once instead of token by token. It still works; it just stops feeling live.

Platforms

Windows, macOS, and Linux. The plugin is desktop-only (isDesktopOnly) because it uses Node APIs for hardware detection — that excludes Obsidian mobile, not any desktop OS.

GPU detection differs by platform, and everything degrades to "VRAM unknown" rather than failing:

On Apple silicon the GPU addresses system memory, so the fit check compares against ~75% of total RAM instead of looking for dedicated VRAM — a 30B model genuinely does fit on a 64 GB Mac, and reporting "no GPU detected" there would be wrong rather than merely unhelpful.

Context window

The context length is read from the model itself — Ollama reports it via /api/show, LM Studio via its native API — and refreshed whenever you switch models. Turn off Detect context window automatically to set the number by hand. Detected values are capped at 131,072 tokens: some models advertise far more than the machine can actually serve, and a gauge scaled to a million tokens would read 0% forever.

That number drives three things: the gauge in the composer, trimming, and the window the server itself uses. When a request would overflow, the oldest messages are dropped and a notice says how many — a model that has quietly forgotten the start of a thread just looks like it got worse, so this is deliberately loud.

It is sent to Ollama as num_ctx. That matters more than it sounds: without it the server runs at its own default no matter what this setting says, so a request carefully trimmed to fit 20k tokens arrives and is truncated to the default anyway — silently, with the dropped tokens never reaching the model. It also means the setting costs video memory, because the KV cache scales with the window. Lowering it is one of the few levers that can make a large model fit on the GPU.

There is no equivalent on the OpenAI-compatible path: LM Studio and friends take the context length from how the model was loaded, not from the request.

GPU layers

Leave this on automatic. Ollama decides how much of a model to place on the GPU and is usually right, and the field is empty by default.

It exists for the case where it is not. On Windows the NVIDIA driver will let a process overcommit video memory and page the excess through system RAM, which thrashes over PCIe and is far slower than a clean CPU offload — the difference between "slow" and "apparently frozen". The symptom is a model performing much worse than the GPU share in the context pane suggests it should. Capping the layers below what fits sidesteps it. 0 runs entirely on the CPU. Ollama only.

Token counts are estimated at ~4 characters per token, and images are counted at a flat 800. Both are approximations meant to drive a gauge, not to match your model's tokenizer exactly.

GPU memory

The same pane reports video memory underneath the token table, because a full context window is only one of the two reasons a reply crawls. It shows card-wide VRAM in use (NVIDIA only — read live from nvidia-smi, never cached) and, for Ollama, how much of each loaded model actually sits on the GPU. The reading requires an external tool; see what this plugin can reach for the exact list and the setting that turns it off.

That last number is the one worth watching. When a model does not fit, Ollama splits it with the CPU rather than failing, and every token then has to cross the CPU-resident layers — so a model 80% on the GPU does not run at 80% speed, it runs at a small fraction of it. That is what a "hang" on a large model almost always is. The fix is a smaller model or quantisation, a lower context window (the KV cache grows with it), or freeing the card and reloading.

What this plugin can reach

Obsidian marks Hephaestus as able to run shell commands and read files outside the vault, which is true and worth being precise about. Both capabilities exist for one feature — reporting your GPU and its memory — and both are behind Detect GPU and video memory in settings. Turn that off and the plugin spawns no process and opens no file outside the vault.

The complete list of what it runs, with the exact arguments:

Both go through execFile with a fixed argument list, so no shell is involved and nothing you type is ever part of a command. Both have timeouts, and a failure — including the tool not being installed — is swallowed and reported as "unknown" rather than raised.

The one file read outside the vault is /sys/class/drm/card{0..3}/device/mem_info_vram_total, a Linux kernel file that reports total video memory for AMD cards. The path is fixed, it is read-only, and nothing is ever written outside the vault.

Two more things Obsidian reports, for completeness. CPU model, core count and RAM come from Node's own os module — in-process, no subprocess. And the plugin lists vault files (getFiles, getMarkdownFiles) to populate the attach-from-vault picker and to run the search_vault tool; it reads a file's contents only through Obsidian's own cachedRead.

Note writing and prompt injection

The model can call write_to_note to append to your open note. Untrusted text reaches the model through web search results and attached files, and either can contain instructions aimed at it. Every write therefore shows a confirmation with the exact text first. The toggle in settings can turn that off; leave it on.

Layout

assets/              Logo, inlined into the bundle at build time
manifest.json        Obsidian plugin manifest
src/main.ts          Plugin, chat view, settings tab, API clients
src/lib.ts           Pure helpers (tokens, protocol translation) — tested
tests/               node --test suites over src/lib.ts
styles.css           Chat styling on Obsidian CSS variables
esbuild.config.mjs   Bundler config
esbuild.lib.mjs      Builds src/lib.ts for the tests

Acknowledgements

Hephaestus is Apache-2.0 licensed (see LICENSE). It stands on the following work:

Notes on what is and is not borrowed:

• No code is copied from llmfit. The System panel was inspired by it after reading its description; the fit heuristic here is far simpler (weights + 20% overhead vs VRAM then RAM), whereas llmfit models quantization, MoE architecture, and memory-bandwidth speed estimates. If you want serious model recommendations, use llmfit itself.
• The OpenAI-compatible client implements the publicly documented chat-completions request/response shape. No OpenAI code is included, and the plugin never contacts OpenAI — the shape is simply what LM Studio, llama.cpp, vLLM, and others expose.
• Web search scrapes html.duckduckgo.com. That is a service, not a dependency: no DuckDuckGo code is bundled, and use is subject to their terms.
• Lucide ships inside Obsidian; this plugin references icon names rather than bundling the icon set.