Koda

by Johannes Kaindl
5
4
3
2
1
Score: 56/100

Description

Koda — agentic vault companion for Obsidian (mirror of git.jkaindl.de)

Reviews

No reviews yet.

Stats

0
stars
3,303
downloads
0
forks
104
days
1
days
1
days
0
total PRs
0
open PRs
0
closed PRs
0
merged PRs
0
total issues
0
open issues
0
closed issues
637
commits

Latest Version

a day ago

Changelog

Changed

  • Two commands drop the plugin name from their names (UI-STANDARD §11). The palette shows "Koda: Open chat" and "Koda: Ask about the selection" instead of "Koda: Open Koda" and "Koda: Ask Koda about the selection" (German: "Chat öffnen", "Zur Markierung fragen"); the ribbon tooltip of the dog icon now reads "Open chat". The command ids stay, so bound hotkeys keep working.

Fixed

  • A very long run of spaces or blank lines no longer freezes Obsidian for seconds. Five places read such a run quadratically: the property overview of list_notes (a note property with tens of thousands of blanks), the remember and skill-description tools (text from the model), the memory file when a line is appended, and the folder placeholder in the rules text. A run of 50,000 characters took about two seconds and the time grew with the square of the length; it is now linear.

  • Credentials that stand directly behind a word or underscore character are now redacted as well. Before a message goes to the model (and into the llm-lab log), key_sk-proj-…, my_AKIA… and x_Bearer … are blanked like the same credential on its own; previously they went out in clear text.

README file from

Github

Koda

🇬🇧 English · 🇩🇪 Deutsch

A companion in your Obsidian vault: a chat sidebar that searches, reads and writes your notes, and asks before it touches anything of yours.

License: AGPL-3.0 Release Obsidian

Koda runs against an OpenAI-compatible model you choose, typically one on your own machine (LM Studio, Ollama), so your notes do not have to leave it. Everything that steers Koda — its instructions, its memory, its skills — is a Markdown note you can read and edit.

Features

  • Chat with your vault. Koda looks things up itself: full-text search, reading notes (including Bases and Canvas files), listing folders with their frontmatter or as a folder tree. Every tool call shows up in the chat, so you see which notes an answer rests on.
  • Writes with your approval. Koda writes freely only in its own folder. Anything else — writing, moving, editing the note you are in — opens a dialog with the text or a line diff first. Deleting and writing skills always ask, and so does replacing the text of the memory note (adding a line to it does not).
  • Working context. Each question can carry what you are looking at: pointers to the open note and selection, or whole notes with their linked neighbours, all open tabs, or the notes that match your question by meaning. The Context tab shows what will go along, and you can leave anything out.
  • Looks at images and makes them. read_image opens an image from your vault: a model that sees images gets the picture, any other model gets the text recognised by Image to Markdown. generate_image asks Local Image Generator for a picture and shows it in the chat, after your approval outside the Koda folder.
  • Bases as context. A view of the type Koda context hands the rows of a base to the conversation as a table, as notes or both; a task can open and take over a base by itself.
  • Tasks for one conversation. Pick the skills, tools and context a job needs before you start: a task is a note, the Abilities tab and / in the input field switch things for this conversation only.
  • Memory and skills as notes. "Remember that …" appends a dated line to Koda/Memory.md. Skills are Markdown notes with standing instructions; Koda can write them for you, with confirmation. A skill with pinned: true in its frontmatter (for example your start routine) is always loaded in full; all others Koda sees by name and description and reads with load_skill when a task matches.
  • Long conversations. Before a conversation overflows the model, Koda compacts it: old tool results first, then a summary of older replies. What you see in the chat and your own messages stay untouched.
  • Tunable for small models. Replace the instructions, switch single tools off, and see the exact prompt the model gets.
  • Works with neighbours, needs none. With Vault Retrieval installed, Koda adds search by meaning; other plugins can lend Koda their tools: they appear in the settings in their own group, each with a switch, and a write that reaches outside the vault says so in the approval dialog. Without them, nothing is missing.

Requirements

  • Obsidian 1.11.4 or newer, desktop or mobile.
  • An OpenAI-compatible chat endpoint with a tool-calling model: a local server (LM Studio, Ollama, …) or a hosted provider with an API key. Models without native tool calling work through a text fallback, less reliably.
  • Claude (Anthropic API) works through the LLM Endpoint Manager: install it, add a Claude endpoint there, and switch on Take endpoints from the LLM Endpoint Manager in Koda's settings. Off by default; without the manager nothing changes.
  • A local server needs CORS switched on (LM Studio: Enable CORS, or lms server start --cors; Ollama: OLLAMA_ORIGINS). The connection test passes without it, the chat does not; Koda names this case when it happens.
  • Optional: Vault Retrieval with an indexed vault, for search by meaning, the Vault context mode and the related_notes tool.
  • Optional: Image to Markdown for reading images with a model that cannot see them, and Local Image Generator for the generate_image tool.

Install

  1. Open Settings → Community plugins → Browse.
  2. Search for Koda, select Install, then Enable.

AnySource Sideloader

With AnySource Sideloader, add https://github.com/johannes-kaindl/koda-agent as a source and install Koda from there. Updates then arrive like any other plugin update.

Manual

Download main.js, manifest.json and styles.css from the latest release into <your vault>/.obsidian/plugins/koda-agent/, then enable Koda under Settings → Community plugins. Each release also ships checksums.sha256 (shasum -a 256 -c checksums.sha256).

Usage

  1. Start your model server, open Settings → Koda, check that the endpoint row (default http://127.0.0.1:1234) shows a green check mark, and pick the model in the endpoint row (Fetch models).
  2. Open the sidebar with the dog icon in the ribbon or the command Open chat.
  3. Ask a question about your vault. The status line says what Koda is doing; each tool call appears in the chat.
  4. Check or change what goes along in the Context tab, or with the dropdown next to Send.
  5. Approve or cancel writes in the dialog that opens before any change outside the Koda folder.

The Getting started guide walks through this in detail.

Configuration

Everything is under Settings → Koda: endpoints (each with its model), the Koda folder (default Koda), how many tool calls one answer may take, compaction, the working context and model control. The settings reference lists every setting with its default and range.

Click the preview for the full-size settings page.

How it works

A question starts an agent loop: the model answers or calls a tool, Koda runs the tool against the vault and hands the result back, up to a limit you set. Writes go through a single rule — free in the Koda folder, a dialog everywhere else — that no tool can bypass. The system prompt is rebuilt for each conversation from Koda's rules, your memory note and your skills. More in How Koda works.

Documentation

Contributing

The canonical repository is git.jkaindl.de/jkaindl/koda-agent; GitHub is its mirror. Issues are welcome on GitHub. Development notes (commands, conventions) are in AGENTS.md in the repository, the structure and the reasons behind it in docs/internal/, the smoke checklist in docs/SMOKE.md.

License

AGPL-3.0-or-later — © 2026 Johannes Kaindl.