README file from
GithubCoDriver
Chat with your Obsidian vault without leaving Obsidian.
CoDriver puts your preferred language models in the right sidebar. Create, improve, and organize notes without giving up control over changes.
What you can do
- Create notes, append content, update text and properties, move files, and send notes to trash.
- Review note patches as a unified diff or side-by-side comparison, then accept, reject, or roll back an applied patch.
- Create and edit Canvas diagrams through reviewed text changes with mandatory JSON Canvas validation.
- Connect OpenAI-compatible providers, Gemini, or Anthropic and switch models from the chat composer.
- Connect external MCP tools and add skills to extend the assistant with reusable instructions and access to your workflows.
- See every vault tool call and control automatic execution per tool.
- Works on desktop and mobile; desktop-only features are marked below.
Add your first provider
- Open Settings -> CoDriver Settings -> Providers.
- Click Add model under LLM providers.
- Choose the Provider type. For OpenAI-compatible services, search Preset by provider or brand name. Selecting a preset fills its endpoint immediately and makes Endpoint read-only. Choose Custom to enter an endpoint; it starts empty. For Alibaba Cloud, choose Custom and enter your regional workspace base URL from its console.
- Enter the API key directly. The field starts masked; use the eye button to show or hide it. Keys stay in Obsidian secret storage. Model loading uses the current input without saving it.
- Click Load models for OpenAI-compatible services, or Test connection for Gemini and Anthropic, to load available models.
- Choose the Default model, then click Save. The optional Model ID field is shown for Custom, Together AI, Fireworks AI, DeepInfra, Z.ai, NVIDIA and Alibaba presets, whose catalog support is unknown or differs from the standard model-list API. It also appears after a failed or empty Load models result. Endpoint edits do not affect its visibility. Manual entry does not confirm chat compatibility. Saved profiles reopen under Custom with their values preserved and an editable endpoint.
Presets include OpenAI, OpenRouter, DeepSeek, Groq, Mistral, Together AI, Ollama, LM Studio, xAI/Grok, Moonshot AI/Kimi, Z.ai/GLM, Alibaba Cloud/Qwen, Fireworks AI, DeepInfra, Hugging Face Inference Providers, NVIDIA API Catalog and SiliconFlow. Model capabilities and optional parameter support depend on the selected model. For Gemini, choose Google Search or Custom tools mode. Google Search mode does not expose MCP tools to the model.
Features
Review and edit notes
Search, read, create, and organize notes with built-in vault tools. Review text and property patches in a unified diff or side-by-side comparison. Writes require your approval unless you enable Auto for that exact tool.

An applied note patch with Roll back available. Rollback requires the live review state and is blocked if the note has changed again.
Support for external MCP tools
Connect HTTP servers or stdio servers (desktop only, with Node subprocess support). HTTP supports Bearer tokens, OAuth 2 (desktop only), Custom headers, and no authentication. Authentication secrets are stored in Obsidian Keychain (1.11.4+; OAuth 2 requires 1.11.5+). Each external tool has separate Enabled and Auto controls for availability and execution without per-call approval.
The experimental Direct API call tool (codriver_api_call) also supports selected Obsidian and installed-plugin APIs. Enable the desired target and configure access and automatic execution individually for each plugin; plugin Auto consent is tied to its installed version. Plugin methods are trusted execution, not sandboxed, and their effects are outside CoDriver's built-in note-patch review guarantees.
Stdio Command and Environment values are ordinary plugin settings, not Keychain entries; avoid putting credentials there.
Skills and commands
Skills provide reusable instructions and local reference material. Their SKILL.md frontmatter supports name and description; invocation (manual, model by default, or always); compatibility (mobile or desktop, with both allowed when omitted); license; custom metadata; and requires.mcp for configured server/tool dependencies. Use {{metadata.<key>}} to insert one explicitly referenced custom metadata value. Requirements do not grant tool permissions.
Markdown links, wikilinks, and embeds can include supporting text files from the skill's folder and subfolders. References are local and non-recursive; skills do not fetch remote files or execute scripts. Removing a model-invoked skill excludes it from that session until manually re-added; always-on skills stay attached.
Commands are prompt templates: {} inserts your input, and {activeNote} inserts a reference to the selected active-note path, not its contents. CoDriver: Run command opens a command input window; CoDriver: Run command instantly starts a selected command that does not require text input. Trigger these Obsidian commands from the command palette or plugins such as QuickAdd, Templater, and Meta Bind to speed up interactions with the assistant.
Work with attachments and audio
Add files with the picker or drag and drop them into the chat input. Supported types include text and code files (such as Markdown, TXT, CSV, JSON, and YAML), DOCX, images (PNG, JPEG, GIF, WebP, BMP, and SVG), and PDF. PDF text is extracted using unpdf. Image support also depends on the selected model.
On desktop, transcribe audio files in .m4a, .mp3, .wav, .webm, .ogg, and .flac formats using the configured audio-transcription provider and model.
Voice input is desktop-only and has its own provider/model selection, independent of audio-file transcription. Record, stop, review the transcribed text, and send it when ready.
Active-note context
Include the current Markdown or Canvas path without automatically sending its contents. Disable automatic selection in Settings or choose a path manually through Add context -> Active note.
Canvas
A built-in skill helps the model generate Canvas JSON. CoDriver's built-in tools validate its structure and references before saving, rejecting invalid JSON even without the skill. Edit the built-in Canvas skill in Settings to describe extensions from plugins such as Advanced Canvas or adapt the instructions to your model. CoDriver's validation still applies; skill edits do not enable unsupported Canvas structures.
Safety and privacy
- Active-note content is never injected automatically.
- Sensitive credentials, including provider API keys and HTTP MCP authentication secrets, are stored in Obsidian Keychain.
- Deleting a provider or MCP server also offers to delete its associated secret.
- Proposed edits are revalidated before application or rollback.
- Configure Max tools, Max calls, Output chars, and Maximum context size to review oversized tool catalogs, automatic call chains, large results, and the combined request context before continuing.
- Request info shows the sent-context breakdown, estimated context tokens, provider-call count, and input/output token usage, with reasoning and cache figures when reported. Use it alongside context limits to monitor token consumption; estimates and incomplete usage are marked.
- Session history stays on your device without saving attachment contents, transcripts, or tool results in it.
- Errors only and Standard diagnostics exclude note content, prompts, provider responses, tool data, transcripts, and secrets.
- Detailed logs exclude reasoning/thinking, prompts, internal provider context, raw attachments/audio and transcription-tool transcripts. An assistant answer or ordinary tool result may still quote sensitive source material.
- Detailed logging requires confirmation each time it is enabled and stays off after restart. The file is refreshed when logging starts in a new CoDriver launch and rotates without archives at 1024 KiB.
Your selected provider still receives the context required for the current request. Review that provider's privacy and retention terms before sending sensitive material.
Installation
From Community plugins
- Open Settings -> Community plugins and turn off Restricted mode if prompted.
- Click Browse and search for CoDriver.
- Open CoDriver, click Install, then Enable.
Manual installation
- Open the latest CoDriver release.
- Download
main.js,manifest.json, andstyles.css. - Copy them into
.obsidian/plugins/codriver/inside your vault. - Open Settings -> Community plugins, turn off Restricted mode, and enable CoDriver.
To update CoDriver, download the latest versions of the same three files, replace the installed copies, and reload Obsidian.
Requirements
- Obsidian
1.5.0or newer. - API access to an OpenAI-compatible provider or local server, Gemini, or Anthropic, with an API key where required. CoDriver connects through provider APIs, not installed coding-agent harnesses or CLI sessions.
- Desktop Obsidian with Node subprocess support for stdio MCP servers.
CoDriver works inside Obsidian. It is not a background automation agent or a CLI coding tool.
[!WARNING] CoDriver is still evolving. For peace of mind, try it in a test vault first and keep regular backups of important notes.
Support
To leave feedback, request a feature, or report a problem, click here.
License
CoDriver is available under the MIT License. See Third-Party Notices for vendored dependencies and their licenses.