Local REST API with MCP

Local REST API with MCP

Give your scripts, browser extensions, and AI agents a direct line into your Obsidian vault via a secure, authenticated REST API.

• Interactive API docs: https://coddingtonbear.github.io/obsidian-local-rest-api/
• Obsidian Community page: https://community.obsidian.md/plugins/obsidian-local-rest-api/

• What you can do
• Quick start
  • REST API
  • MCP clients
    • Claude Code
    • Claude Desktop
    • Cursor
    • Other clients
• API overview
  • The configuration directory is off-limits
  • Failed authentication is throttled
  • Conditional writes
  • Browser clients and response headers
• Patching notes
  • Raw-content mode
• Targeting specific sections
• Searching
• Event streams
• Server state
• MCP (Model Context Protocol)
  • Protocol revisions
  • Connecting a client
  • Available tools
  • Binary files and attachments
  • Signed URLs
  • Available resources
• API Extensions
  • Typed extension API
  • Registering again after the host reloads
  • MCP tools, resources, and prompts
  • Documenting your routes
  • Sub-resources under a note
  • Extension events
  • Extension state
  • Known extensions
• Contributing
• Credits

What you can do

Access your vault through the REST API or the built-in MCP server — both interfaces expose the same core capabilities, so scripts, browser extensions, and AI agents all speak the same language.

• Read, create, update, or delete notes — full CRUD on any file in your vault, including binary files
• Surgically patch specific sections — target a heading, block reference, or frontmatter key and append, prepend, replace, delete, or move just that section without touching the rest of the file
• Search your vault — simple full-text search or structured JsonLogic queries against note metadata (frontmatter, tags, path, content)
• Follow vault events — subscribe to Obsidian events (a note created, its frontmatter changed, a file opened) as a filtered Server-Sent Events stream
• Access the active file — read or write whatever note is currently open in Obsidian
• List and execute commands — trigger any Obsidian command as if you'd used the command palette
• Query tags — list all tags across your vault with usage counts
• Open files in Obsidian — tell Obsidian to open a specific note in its UI
• Extend the API — other plugins can register their own routes via the API extension interface

All requests are served over HTTPS with a locally generated certificate and gated behind API key authentication.

Quick start

After installing and enabling the plugin, open Settings → Local REST API to find your API key and certificate.

REST API

# Check the server is running (no auth required)
curl -k https://127.0.0.1:27124/

# List files at the root of your vault
curl -k -H "Authorization: Bearer <your-api-key>" \
  https://127.0.0.1:27124/vault/

# Read a note
curl -k -H "Authorization: Bearer <your-api-key>" \
  https://127.0.0.1:27124/vault/path/to/note.md

# Read a specific heading (URL-embedded target)
curl -k -H "Authorization: Bearer <your-api-key>" \
  https://127.0.0.1:27124/vault/path/to/note.md/heading/My%20Section

# Append a line to a specific heading (PATCH with a JSON instruction)
curl -k -X PATCH \
  -H "Authorization: Bearer <your-api-key>" \
  -H "Content-Type: application/json" \
  --data '{"targetType":"heading","target":["My Section"],"operation":"append","content":"New line of content"}' \
  https://127.0.0.1:27124/vault/path/to/note.md

To avoid certificate warnings, you can download the plugin's certificate authority from https://127.0.0.1:27124/obsidian-local-rest-api.crt and trust it in your OS or browser, or point your HTTP client at it directly (for example curl --cacert obsidian-local-rest-api.crt ...). The plugin generates its own certificate authority on first run and serves a server certificate signed by it, so the download is a CA certificate rather than the server certificate itself. That CA is name-constrained: it can only vouch for 127.0.0.1, localhost, your configured binding host, and the hostnames you list under Subject alternative names, so trusting it does not let it (or anyone who obtains its key) impersonate other sites.

MCP clients

The MCP server runs at https://127.0.0.1:27124/mcp/ and requires that you provide your bearer token for authentication via an Authorization header (i.e. Authorization: Bearer <your-api-key>). Because the plugin uses a locally generated certificate authority, you may need to either trust that certificate in your OS/client, or use the plain HTTP endpoint at http://127.0.0.1:27123/mcp/ (enable it under Settings → Local REST API → Enable HTTP server).

Claude Code

Claude Code has native HTTP MCP support. The quickest way to add the server is via the CLI:

claude mcp add --transport http obsidian https://127.0.0.1:27124/mcp/ \
  --header "Authorization: Bearer <your-api-key>"

Or add it manually to .mcp.json in your project root (project-scoped) or configure it user-wide via claude mcp add --scope user:

{
  "mcpServers": {
    "obsidian": {
      "type": "http",
      "url": "https://127.0.0.1:27124/mcp/",
      "headers": {
        "Authorization": "Bearer <your-api-key>"
      }
    }
  }
}

Claude Desktop

Claude Desktop does not natively support remote HTTP MCP servers, but you can bridge it with mcp-remote (requires Node.js). Add the following to claude_desktop_config.json:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "obsidian": {
      "command": "npx",
      "args": [
        "mcp-remote@latest",
        "https://127.0.0.1:27124/mcp/",
        "--header",
        "Authorization: Bearer <your-api-key>"
      ]
    }
  }
}

Restart Claude Desktop after saving the file.

Cursor

Cursor supports the Streamable HTTP MCP transport. Add the following to ~/.cursor/mcp.json (global) or .cursor/mcp.json (project-specific):

{
  "mcpServers": {
    "obsidian": {
      "url": "https://127.0.0.1:27124/mcp/",
      "headers": {
        "Authorization": "Bearer <your-api-key>"
      }
    }
  }
}

Other clients

Any MCP client that supports the Streamable HTTP transport can connect to https://127.0.0.1:27124/mcp/ with an Authorization: Bearer <your-api-key> header. Consult your client's documentation for the exact configuration format.

API overview

For full request/response details, see the interactive docs.

The configuration directory is off-limits

The API refuses to read or write any file inside Obsidian's configuration directory (app.vault.configDir, normally .obsidian) — for both REST and MCP, reads as well as writes. A request for such a path is rejected with 403 (error code 40321) at the REST layer or a path error from an MCP tool.

That directory holds plugin code and each plugin's data.json — including this plugin's own, where your API key is stored. Writing into it is effectively remote code execution, since Obsidian runs an enabled plugin's main.js; reading from it leaks those secrets. Blocking it keeps a credential documented as "vault file access" from silently granting more (see GHSA-66m9-r757-qvq7).

The check is against where a path lands on disk, not just how it is spelled: a Windows 8.3 short name such as OBSIDI~1, a differently cased spelling on a case-insensitive filesystem, or a symlink that points into the configuration directory is refused the same way. Search and the tag listing skip any indexed note that lives there too, so a symlinked folder cannot hand the contents out by another route.

If you deliberately manage your Obsidian configuration through the API, turn on Settings → Local REST API → Advanced settings → Allow access to the configuration directory. It is off by default, and turning it on grants every holder of your API key that access.

Failed authentication is throttled

A source that presents a wrong credential -- an incorrect API key, or an invalid, expired or spent signed URL -- ten times within a minute has its further wrong credentials refused with 429 (error code 42901) and a Retry-After header until the minute is up. This applies to every route the plugin authenticates, including /mcp/ and GET / (routes an extension plugin publishes without authentication are not behind it): the root route answers 200 with authenticated: false to a wrong key, so it would otherwise be the cheapest place to guess from, and a wrong key sent there counts like a wrong key anywhere else.

Nothing else is slowed down. A request carrying the correct key or a valid signed URL is never counted or delayed, so MCP clients and scripts that make many requests in quick bursts are unaffected, and so is a client with the right key while another process on the same machine is being refused. A request that presents no credential at all is not counted either -- it has made no guess -- and is answered 401, or served on the routes that need no authentication, however many of them arrive. The counter is kept in memory per source address and is cleared by a plugin reload.

Conditional writes

Two clients that read the same note and then each write their edited copy back would otherwise lose the first write without a word. Every write to a file can carry a precondition instead, so it fails rather than overwriting a change it never saw:

# Read the note; the ETag response header is its current version
curl -i -H "Authorization: Bearer $KEY" https://127.0.0.1:27124/vault/notes/todo.md
# => ETag: "a1b2c3"

# Write it back only if nobody has changed it since
curl -X PUT -H "Authorization: Bearer $KEY" -H 'If-Match: "a1b2c3"' \
     -H "Content-Type: text/markdown" --data-binary @todo.md \
     https://127.0.0.1:27124/vault/notes/todo.md
# => 204, with the new version in ETag -- or 412 if the note changed, and nothing written

# Create a note only if it does not exist yet
curl -X PUT -H "Authorization: Bearer $KEY" -H "If-None-Match: *" \
     -H "Content-Type: text/markdown" --data-binary @new.md \
     https://127.0.0.1:27124/vault/notes/new.md

If-Match works on every write: PUT, POST, PATCH and DELETE on /vault/{path} and /active/, MOVE and COPY on /vault/{path}, and writes aimed at a heading, block or frontmatter field. It takes the ETag from a GET, the version from a document map or note JSON — all three are the same token — or the ETag a previous write answered with. A list matches if any entry does, and * matches any file that exists. A failed precondition is 412 (error code 41200), and the message names the file's current version. Writes made through the API are queued per file, so two clients holding the same version cannot both pass the check.

The ETag sent with a note's HTML rendering or its note JSON is not this token, because those also depend on other files; read note JSON's version field instead. MCP's write tools take the same precondition as an ifMatch argument (and vault_write/vault_append an ifNoneMatch: "*"), vault_read returns the version, and the write tools return the new one.

Browser clients and response headers

Several endpoints answer in a response header rather than in the body: ETag carries a file's version for conditional writes, Content-Location tells you which file a targeted or /active/ request actually resolved to, Markdown-Patch-Warnings reports what a PATCH had to work around, Deprecation warns that a format is sunsetting, and Mcp-Session-Id carries the session for a sessionful MCP connection.

Browsers hide response headers from JavaScript unless the server opts them in, so the API sends Access-Control-Expose-Headers: * and all of them are readable with response.headers.get(...). Safari honours the wildcard from 15.4 onward; older browsers see only the CORS-safelisted headers. Requests made with credentials: "include" are not supported — the API authenticates with a bearer token and sends Access-Control-Allow-Origin: *, which browsers reject for credentialed requests.

Patching notes

The PATCH method is one of the most useful features of this API. It lets you make targeted edits without rewriting entire files.

Send a JSON instruction: an operation (replace, prepend, append, or delete) applied to a scope (content, marker, markerAndContent, or parent) of a target — a heading (addressed as an array of heading texts from the top level down), a block reference, or a frontmatter key. The payload rides in content (a string), value (JSON, for frontmatter values), or destination (a heading move):

# Replace the value of a frontmatter field
curl -k -X PATCH \
  -H "Authorization: Bearer <your-api-key>" \
  -H "Content-Type: application/json" \
  --data '{"targetType":"frontmatter","target":"status","operation":"replace","value":"done"}' \
  https://127.0.0.1:27124/vault/path/to/note.md

Heading levels inside a content string are relative to the target (a leading # becomes a direct child). Advisory warnings (e.g. a heading rebased past level 6) come back as percent-encoded JSON in the Markdown-Patch-Warnings response header — decode with decodeURIComponent before parsing. Pass ifMatch (the version from a document map) for optimistic concurrency, or send an If-Match header as with any other conditional write.

Note: Whitespace is library-owned — your content is reduced to trimmed, canonical form (leading and trailing blank lines are meaningless), and the API itself supplies the blank line wherever inserted content faces body text, so an append or prepend always lands as its own block and never merges into an existing paragraph. Heading lines, existing blank lines, and each document's spacing style are preserved as-is. See the interactive docs for worked examples.

To continue an existing block instead of starting a new one — say, extending a list — add within to a heading instruction: an index selecting one of the section's top-level body blocks (0-based in document order, negative counting from the end, so -1 is the last block). A within edit splices literally, so you own the joint:

# Add an item to the last list under "Log" (the leading \n continues the block)
curl -k -X PATCH \
  -H "Authorization: Bearer <your-api-key>" \
  -H "Content-Type: application/json" \
  --data '{"targetType":"heading","target":["Log"],"within":-1,"operation":"append","content":"\n- new item"}' \
  https://127.0.0.1:27124/vault/path/to/note.md

With markerAndContent scope, prepend/append instead insert a new block beside the indexed one. Indices are positional, so read the document map first and pair the edit with ifMatch.

Raw-content mode

If your client templates markdown into the request body (Shortcuts, Tasker, curl from a template), JSON-escaping that content into an instruction is fragile. Raw-content mode moves the instruction's fields out of the body — target in the URL (or in Target-Type/Target headers with an explicit Markdown-Patch-Version: 2), operation and options in headers — and the body is the raw payload, no JSON escaping required:

# Append a templated line under a heading — no JSON escaping anywhere
curl -k -X PATCH \
  -H "Authorization: Bearer <your-api-key>" \
  -H "Operation: append" \
  -H "Content-Type: text/markdown" \
  --data "- $TEMPLATED_CONTENT" \
  https://127.0.0.1:27124/vault/notes/daily.md/heading/Log

A text/* body is the content carrier, an application/json body the value carrier, and no body at all carries nothing (a delete, or a move via a Destination header). Target-Scope, Within (the instruction's within index as a plain integer, e.g. -1), Create-Target-If-Missing, Reject-If-Content-Preexists, and If-Match headers round out the instruction. See the interactive docs for the header encodings and the full details.

Already using the older header-driven PATCH format? It spread the instruction across request headers instead of a JSON body, and is deprecated and will be removed in 6.0. It still works — send Markdown-Patch-Version: 1 to opt back into it (the same header also selects the legacy ::-joined document map on GET), and responses served by it carry a Deprecation: true; sunset-version="6.0" header. To upgrade, drop that header and move each header into the JSON body; the interactive docs have the field-by-field mapping table. If your content contains headings, adjust their levels too: 1.x wrote them literally, and 2.x treats them as relative to the target (see the migration guide there).

See the interactive docs for the full instruction schema and options.

Targeting specific sections

You can read or write a specific part of a note — a heading, block reference, or frontmatter field — without fetching or replacing the whole file. This works on GET, PUT, POST, and PATCH requests (for PATCH this is raw-content mode — add an Operation header).

Append /<target-type>/<target> after the filename. Each nested heading level is its own path segment, so a heading whose text contains :: needs no escaping:

# Read the content under a specific heading
curl -k -H "Authorization: Bearer <your-api-key>" \
  https://127.0.0.1:27124/vault/path/to/note.md/heading/My%20Section

# Read a nested heading (one path segment per level)
curl -k -H "Authorization: Bearer <your-api-key>" \
  https://127.0.0.1:27124/vault/path/to/note.md/heading/Work/Meetings

# Read a frontmatter field
curl -k -H "Authorization: Bearer <your-api-key>" \
  https://127.0.0.1:27124/vault/path/to/note.md/frontmatter/status

# Replace the content of a heading via PUT (heading levels are normalized for you)
curl -k -X PUT \
  -H "Authorization: Bearer <your-api-key>" \
  -H "Content-Type: text/markdown" \
  --data "Updated content" \
  https://127.0.0.1:27124/vault/path/to/note.md/heading/My%20Section

# Append to a heading via POST
curl -k -X POST \
  -H "Authorization: Bearer <your-api-key>" \
  -H "Content-Type: text/markdown" \
  --data "Appended content" \
  https://127.0.0.1:27124/vault/path/to/note.md/heading/My%20Section

Supported target types: heading, block, frontmatter.

A targeted URL is ambiguous on its face — /vault/notes/log.md/heading/Today could name the Today section of notes/log.md or a file literally called notes/log.md/heading/Today. The server walks backwards down the path until it finds a real file and reports which one it settled on in a Content-Location response header, with each path component percent-encoded on its own (non-ASCII characters, and reserved characters like #, ? and ,) so it can be pasted straight back into a request URL. A request whose URL names the file outright gets no such header.

On a GET, a Target-Scope header selects which part of the target comes back, mirroring the PATCH scopes: content (the default), marker (the label — a heading's raw text, a block's bare id, a frontmatter key), or markerAndContent (the whole node, in exactly the shape a PATCH replace at that scope consumes — a heading subtree reads back with its own line as # Title, levels relative to its parent):

# Read a whole section — heading line included — ready to edit and write back
curl -k -H "Authorization: Bearer <your-api-key>" \
  -H "Target-Scope: markerAndContent" \
  https://127.0.0.1:27124/vault/path/to/note.md/heading/My%20Section

Deprecated: header-based targeting. Earlier releases targeted a section with Target-Type, Target, and Target-Delimiter headers (plus Target-Scope/Trim-Target-Whitespace). That form is deprecated and will be removed in 6.0; it is only processed when you also send Markdown-Patch-Version: 1 (responses then carry a Deprecation header). Without it, supplying those targeting headers is rejected with 400. Supplying both URL-path targeting and the header form on one request returns 422 Unprocessable Entity.

Searching

POST /search/simple/?query=your+terms runs Obsidian's built-in fuzzy search and returns matching filenames with scored context snippets. Obsidian matches each word of the query separately, so a query containing a common word can match thousands of times; &maxResults= caps how many files come back (keeping the highest-scoring ones) and &maxMatchesPerFile= caps the matches within each. Each result's totalMatches and the Total-Results response header say how much was left out.

POST /search/ accepts a JsonLogic expression (content type application/vnd.olrapi.jsonlogic+json) and evaluates it against each note's metadata (frontmatter, tags, path, content).

Event streams

You can follow what happens in the vault as a Server-Sent Events stream. There are two steps. First, register a subscription to one Obsidian event, with an optional JsonLogic filter. Then open the URL that comes back:

# 1. Subscribe to notes under journal/ being modified
curl -X POST -H "Authorization: Bearer <your-api-key>" \
  -H "Content-Type: application/vnd.olrapi.jsonlogic+json" \
  -d '{"glob": ["journal/*", {"var": "path"}]}' \
  https://127.0.0.1:27124/events/vault/modify/
# => {"id": "…", "url": "https://127.0.0.1:27124/events/vault/modify/…/?sig=…&exp=…&n=…", …}

# 2. Follow the stream; with signed URLs on, the URL needs no API key
curl -N "<url>"

It takes two steps because a browser's EventSource can only make GET requests, which have no body to carry a filter.

The events are Obsidian's own, and only these can be streamed:

Each event is serialized by code written for it. That code decides exactly what is sent: the path, the file's NoteJson (the same shape /search/ evaluates), and a few event-specific fields such as oldPath on a rename. Note content is sent only when the filter reads file.content. Events whose payloads are keystrokes, clipboard data, or UI objects (editor-change, quick-preview, editor-paste, the menu events, …) can't be streamed. To react to frontmatter changes, use metadataCache changed: vault modify fires before Obsidian has re-read the file's metadata.

Each message's id is <epoch>-<counter>. A new epoch, or a gap in the counter, means events were missed. Nothing is replayed. A stream URL expires after the signed-URL lifetime (or ?ttl=<seconds>), but a stream opened before then stays open. At most 16 streams can be open at once. Anyone holding a signed stream URL sees the paths and metadata of every event its filter matches, so treat it like the notes themselves. See the API docs for the full message format.

Server state

GET / answers without an API key, but an authenticated request gets more: certificateInfo, the list of installed API extensions, and a state section.

state holds observations, by namespace, that a client reads to decide whether to proceed. The server attaches no verdict. The plugin's own namespace is metadataCache:

{
  "status": "OK",
  "authenticated": true,
  "state": {
    "metadataCache": {
      "listeningSince": "2026-10-03T14:02:11.408Z",
      "lastResolvedAt": null,
      "lastActivityAt": "2026-10-03T14:02:13.951Z"
    }
  }
}

It exists because a note's links, backlinks, and unresolvedLinks come from Obsidian's vault-wide link graph, and while Obsidian is still indexing after startup they can be incomplete with no sign in the note itself. Obsidian gives no documented signal for the end of that indexing, so the plugin reports what it can observe: when it started listening (listeningSince, which is when it loaded), when it last heard the cache's resolved event (lastResolvedAt, fired each time Obsidian's resolver queue drains, or null if never since loading), and when it last heard any indexing activity at all (lastActivityAt: changed, resolve, or resolved). Indexing in progress looks like recent activity; done looks like silence. A plugin enabled into an already-indexed vault hears nothing, which is why listeningSince is there: "never resolved, but listening for a while now" is as good as done. A workable rule is to wait while lastResolvedAt is null and listeningSince is recent, or while lastActivityAt is recent, with "recent" set to whatever your vault's indexing takes. A client that would rather not poll can watch the metadataCache event stream instead, where resolve fires per file and resolved on each drain.

Every other key of state is an extension plugin's id, holding whatever that extension publishes (see Extension state), or null when it could not be read in time. The budget is Settings → Local REST API → Advanced settings → Extension state read budget, 100 ms by default; extensions are read concurrently, so it bounds the whole wait. MCP clients get the same document from the obsidian://local-rest-api/status resource.

MCP (Model Context Protocol)

[!NOTE] Several third-party MCP servers for Obsidian exist, but they are no longer necessary — this plugin ships a built-in MCP server that runs inside Obsidian and has direct access to your vault's live metadata, active file, and command palette. If you are currently using a third-party server, switching to this one is likely to give you better results.

The plugin includes a built-in MCP server at /mcp/ so AI agents and MCP-compatible clients can interact with your vault without hand-crafting HTTP requests.

Transport: Streamable HTTP — API key authentication required.

Protocol revisions

The endpoint serves the 2026-07-28 revision plus the sessionful revisions from 2024-10-07 through 2025-11-25, choosing per request, so clients on either can share it.

The 2026-07-28 revision is stateless: there is no initialize handshake and no session, so the plugin neither issues nor reads the Mcp-Session-Id header. Each request carries its own protocol version and client identity in params._meta, repeats them in the MCP-Protocol-Version, Mcp-Method, and Mcp-Name headers, and is answered on its own. Clients can call server/discover to learn the supported revisions and capabilities up front.

Clients that open with an initialize request are served the sessionful revision they negotiate: the handshake returns an Mcp-Session-Id, GET /mcp/ opens that session's notification stream, and DELETE /mcp/ ends it. Sessions exist only on this path, and they are what keeps the handshake's listChanged capabilities honest: when another plugin registers or removes an MCP tool, every live session is notified, while 2026-07-28 clients hear about it on a subscriptions/listen stream.

Connecting a client

Connect your MCP client to https://127.0.0.1:27124/mcp/. Authentication uses a bearer token — find your API key under Settings → Local REST API, then pass it as:

Authorization: Bearer <your-api-key>

The exact config syntax varies by client; see the Quick start examples above or consult your client's documentation for Streamable HTTP remote MCP servers.

[!WARNING] To connect to the MCP server securely, your client must trust the plugin's locally generated certificate authority. You can download and trust it from https://127.0.0.1:27124/obsidian-local-rest-api.crt, or configure your client to skip TLS verification for 127.0.0.1.

If trusting a locally generated certificate is not possible in your environment, you can connect insecurely using http://127.0.0.1:27123/mcp/ instead of https://127.0.0.1:27124/mcp/ if you have enabled the HTTP endpoint under Settings → Local REST API → Enable HTTP server.

Available tools

search_simple returns at most the 20 highest-scoring files and the first 10 matches in each unless the call sets maxResults or maxMatchesPerFile. Obsidian's simple search matches each word of a query separately, so an unbounded search containing a common word can return megabytes, and an MCP result usually goes straight into the model's context. When anything was left out, a second text block beginning [Search results limited: says how much and how to see more. Its contextLength is at most 500 characters. POST /search/simple/ takes the same two parameters but has no limit by default, and does not bound contextLength.

Binary files and attachments

The REST API has always handled binary content: GET /vault/<path> returns raw bytes with a Content-Type derived from the file extension, and PUT /vault/<path> accepts a body of any content type and stores it byte-for-byte. Neither has a practical size limit.

MCP tools are a different story, because a tool's arguments and results pass through the model. vault_read and vault_write are text tools — they decode and encode UTF-8, which is lossy for anything that is not text — so vault_read refuses a file whose bytes are not valid UTF-8, and vault_write and vault_append refuse a path whose extension names an image (other than SVG, which is text), audio, video, font, PDF, or archive type. Reading an attachment as text and writing the result back is the mistake that destroys attachments, and both halves of it are now refused.

vault_read_binary is the tool for attachments, and what it returns depends on the file:

• Raster images come back as an MCP image content block, downscaled to fit 1568px on the long side, plus a small text block with the file's path, MIME type, size, and dimensions. The model can actually look at the picture, and is billed for its pixels rather than its bytes: a multi-megabyte photo costs a couple of thousand tokens. An image that is still over 512 KiB once downscaled -- a large image that was already inside 1568px, so nothing was resized -- comes back as a download link instead, the same as any other oversized file, or is refused with a pointer at the REST endpoint when signed URLs are off and there is no link to give.
• SVGs come back unchanged, as their source text in a resource block. A vector drawing is XML the model can read directly, so nothing is rasterized or resized.
• Everything else comes back as a resource_link to a signed download URL when signed URLs are enabled (below), so the bytes never enter the conversation. When they are not enabled, a file under 512 KiB is embedded as a resource block with base64 bytes, and a larger one is refused with a pointer at the REST endpoint.

An as argument overrides the default: as: "bytes" embeds the raw bytes (under 512 KiB), as: "link" returns a signed link and never reads the file.

The 512 KiB ceiling is not only about token cost. A tool result carrying roughly a megabyte or more of base64 crashes Obsidian's Electron renderer outright, taking the plugin's HTTP server down with it, so the cap is set well below the point where that was observed.

There is no upload tool that carries bytes through the model — emitting base64 as output tokens is impractical beyond a few kilobytes. The agent's host has the file on disk; it uploads it with a PUT to the REST API, either with the API key or with a signed upload URL.

Signed URLs

Signed URLs let an agent hand a file to something that is not the MCP client — a browser tab, an <img> tag, a curl in a shell — without also handing over the API key. They are on by default; turn them off under Settings → Local REST API → Advanced settings → Enable signed URLs, and set their lifetime there (default 300 seconds).

While they are on, these MCP tools mint them:

• vault_get_download_url returns a resource_link to GET /vault/<path>?sig=…&exp=…&n=…, plus a markdown link for clients that only render text. The link is valid until it expires and can be used repeatedly. Add &download=1 to have the browser save the file instead of showing it.
• vault_get_upload_url returns a PUT /vault/<path>?sig=…&exp=…&n=… URL and a ready-to-run curl command, with the filename quoted for a POSIX shell so a name containing $, backticks or spaces cannot be expanded when the command is pasted. The link is consumed by the first request that succeeds — claimed at authorization rather than at completion, so concurrent redemptions cannot all pass, and released again if the request does not end in a 2xx. The Content-Type is informational on a signed upload: the bytes are stored exactly as sent, whatever type is declared, because a signed URL authorizes a whole-file write of that content. Without it, a .json destination was parsed and re-serialized, so a pretty-printed file lost its whitespace and trailing newline while still answering 204.
• vault_read_binary uses download links for non-image files, and for anything when called with as: "link".
• events_get_listener_url registers an event stream subscription and returns its GET /events/<emitter>/<event>/<id>/?sig=…&exp=…&n=… URL, plus a curl -N command. POST /events/<emitter>/<event>/ returns the same kind of URL.

A signed URL authorizes a whole-file write to the path it names, and only that: a request that also carries Target-Type/Target headers, or whose path continues into /heading, /block or /frontmatter, is refused rather than quietly becoming a targeted edit of a document the link never named. Use the API key for targeted writes.

How they work: the signature is an HMAC over the method, the normalized vault path, the expiry, and a random per-link nonce carried as n, under a secret generated fresh every time the plugin loads and kept only in memory. A link is therefore good for one file, one method, one window of time, and never survives an Obsidian restart. The nonce is what makes each mint its own link: exp has one-second granularity, so without it two links minted for the same file in the same second were byte-identical -- and since a spent upload link is remembered by its signature, re-minting straight after an upload handed back the link that had just been consumed. The host is not part of the signature, so the same link works whichever hostname on the certificate the client uses. Anyone holding a link can do what it names until it expires, so treat one as you would the file itself.

Two practical notes: whether a chat client renders a linked image inline is up to the client, and most do not today, but clicking through always works; and a link on the HTTPS port needs the plugin's certificate to be trusted by whatever opens it — the plain-HTTP port avoids that.

Available resources

API Extensions

Other plugins can register their own authenticated routes, public routes, MCP tools, streamable events, and state against this plugin's server. See Adding your own API Routes via an Extension for a walkthrough.

Public routes (addPublicRoute) are answered before the API key is checked, so they can't sit under a prefix the plugin serves its own routes from: /vault/, /active/, /search/, /commands/, /events/, /mcp/, /open/, and /tags/, in any letter case, along with /, the OpenAPI documents, and the certificate. A path whose first segment is a pattern (/:name/, /*) is refused for the same reason. addPublicRoute throws when you register one of these, so start public routes with a literal segment of your own, such as your plugin's id. Authenticated routes (addRoute) and vault sub-resources are unaffected.

Typed extension API

Install this package as a development dependency to get getAPI and the types for everything it returns:

npm install --save-dev obsidian-local-rest-api

This package declares obsidian, zod, and @types/express as peer dependencies, because its types refer to all three — addRoute returns express's IRoute, and addMcpTool takes zod schemas. npm installs peers for you; if you pin them yourself, keep them resolvable. Without them, TypeScript quietly widens those positions to any instead of reporting an error, so a project that suppresses the missing-types diagnostic gets no warning that it has lost type checking exactly where it matters most.

import { getAPI, type LocalRestApiPublicApi } from "obsidian-local-rest-api";

const api: LocalRestApiPublicApi | undefined = getAPI(this.app, this.manifest, 2);

The package entry point is a small standalone module — it resolves the running host plugin out of Obsidian's plugin registry rather than pulling the plugin bundle into your build. Passing an extension API version (2 above) makes getAPI throw ApiVersionUnsupportedError when the installed host is older than the surface you need; omit it to accept whatever is installed and feature-detect yourself. getAPI returns undefined when the plugin isn't installed or hasn't loaded yet.

publicApi.d.ts is generated from src/publicApi.ts, which the implementation is compile-time-checked against, so the published types cannot drift from what the plugin actually offers.

Registering again after the host reloads

Everything an extension registers belongs to the running instance of this plugin. When this plugin is disabled and re-enabled, updated, or reloaded during development, it starts with a new server that knows nothing about your routes, tools, or events, and the handle getAPI gave you earlier now points at the instance that was unloaded. Your routes answer 404 from then on, with no error to explain why.

Each time this plugin finishes loading, it triggers obsidian-local-rest-api:loaded on the workspace. Register in a method of your own, call it from onload, and call it again whenever that event fires:

async onload() {
  this.registerWithLocalRestApi();
  this.registerEvent(
    // Not in obsidian's typings; declare it on Workspace, or cast as here.
    (this.app.workspace as Events).on("obsidian-local-rest-api:loaded", () =>
      this.registerWithLocalRestApi()
    )
  );
}

registerWithLocalRestApi() {
  // A fresh handle every time: one kept from before a reload is stale.
  this.api = getAPI(this.app, this.manifest, 3);
  if (!this.api) return; // Not installed or not loaded yet; the event will call again.
  this.api.addRoute("/my-plugin/status").get((req, res) => { /* ... */ });
}

Calling it from onload covers the case where this plugin loaded before yours, and the event covers every load after that, including this plugin loading after yours at startup. Avoid guarding the method with if (this.api) return: that keeps the stale handle and skips registering on exactly the occasions it's needed. The sample extension shows the complete pattern, including handling a host that is too old and cleaning up in onunload.

MCP tools, resources, and prompts

addMcpTool(name, description, schema, callback) sends whatever your callback returns back to the client as a single block of JSON text. From extension API version 3 you can instead pass a definition object, and the callback returns the complete MCP result, which reaches the client unchanged. Use it when you need images, structuredContent checked against an outputSchema, or an isError result that tells the model a call failed in a way it can recover from.

Everything in this section needs version 3, so ask for it when you call getAPI. The types describe the whole interface whichever version you pass, so an extension that asks for 2 still compiles against these methods, and then finds them missing at runtime on an older host:

const api = getAPI(this.app, this.manifest, 3);

api?.addMcpTool({
  name: "comments_count",
  description: "Count the comments on a note",
  inputSchema: { path: z.string() },
  outputSchema: { count: z.number() },
  callback: async ({ path }) => {
    const count = await countComments(path as string);
    return {
      content: [{ type: "text", text: `${count} comments` }],
      structuredContent: { count },
    };
  },
});

Version 3 also lets an extension expose things that aren't tools:

• addMcpResource({ name, uri, read }) adds a resource at a fixed URI.
• addMcpResourceTemplate({ name, uriTemplate, read, list? }) adds a family of resources addressed by an RFC 6570 template such as tandem://comments/{path}. read gets the matched variables. list is optional, and when you provide it, its resources appear in resources/list.
• addMcpPrompt({ name, argsSchema?, callback }) adds a prompt. MCP passes prompt arguments as strings.

Clients that are already connected are notified when these lists change, and unregister() removes everything the handle registered.

Documenting your routes

The plugin can't see what an extension's routes accept or return, so they don't appear in the OpenAPI spec until the extension describes them. addOpenApiDescription (extension API version 3) takes the paths, components, and tags your routes need, in the same shape as the matching parts of an OpenAPI document, and merges them into the spec served at /openapi.yaml, /openapi.json, and the MCP openapi-spec resource:

Request version 3 from getAPI so an older host fails loudly instead of lacking the method:

const api = getAPI(this.app, this.manifest, 3);
api.addRoute("/widgets/:id/").get(handler);
api.addOpenApiDescription({
  paths: {
    "/widgets/{id}/": {
      get: {
        tags: ["Widgets"],
        summary: "Return one widget.",
        parameters: [{ name: "id", in: "path", required: true, schema: { type: "string" } }],
        responses: { "200": { description: "The widget." } },
      },
    },
  },
  tags: [{ name: "Widgets", description: "Routes added by the Widgets plugin." }],
});

Write path parameters the OpenAPI way ({id}), not express's (:id). Each path you contribute is published with an x-obsidian-extension field set to your plugin ID. A path, component, or tag that the plugin or another extension already declares makes the call throw without publishing anything, and unregister() removes your description along with your routes.

Sub-resources under a note

Routes added with addRoute can't live under /vault/, because the plugin's own /vault/* handler claims those paths first. From extension API version 3, addVaultSubresource(name) lets an extension serve routes under any note instead:

import type { VaultSubresourceRequest } from "obsidian-local-rest-api";

const comments = api.addVaultSubresource("comments");
comments.get("/", (req, res) => {
  const { vaultFile } = req as VaultSubresourceRequest;
  res.json(listComments(vaultFile));
});
comments.get("/:id", (req, res) => { /* ... */ });

GET /vault/Notes/draft.md/comments/a1f3 then reaches that router as GET /a1f3, with the note attached as req.vaultFile; /active/comments/a1f3 does the same for the active file. The plugin resolves the note before your router runs, so it only ever sees notes that exist, and a request your router doesn't answer continues to the plugin's own handlers. Requests need the API key; signed URLs never reach a sub-resource.

A %2F in the URL is a literal slash inside one segment, which Express's own route matching can't tell apart from a separator. req.vaultSubresourceSegments holds the segments after the name, each decoded on its own, for when that matters.

A name is one path segment. heading, block and frontmatter are reserved, and each name can only be registered by one extension at a time.

Extension events

From extension API version 3, an extension can add its own events to the event streams. Each one is streamed under the extension's plugin id as the emitter:

const api = getAPI(this.app, this.manifest, 3);

// Your plugin's own Events instance; call this.events.trigger("task-completed", ...)
// wherever the event happens.
this.events = new Events();

api.addStreamableEvent("task-completed", {
  source: this.events,
  serialize: (file, line) => ({ path: (file as TFile).path, line }),
});
// Now available at POST /events/<your plugin id>/task-completed/

Your serializer decides everything a stream sends. The host adds emitter and event and sends nothing else, so return only what someone holding a stream URL should see. Return null to skip an occurrence. unregister() closes any open streams for your events.

Extension state

From extension API version 4, an extension can publish its own namespace in the state section of GET /, under its plugin id, for anything a client would read to decide whether to proceed: whether an index is built, when a sync last ran, how much work is queued.

const api = getAPI(this.app, this.manifest, 4);

api.addState({
  description: "Whether the semantic index is ready to query, and how far along a rebuild is.",
  schema: {
    type: "object",
    required: ["ready", "pending"],
    properties: {
      ready: { type: "boolean" },
      pending: { type: "integer", description: "Notes not yet indexed." },
      lastRunAt: { type: ["string", "null"], format: "date-time" },
    },
  },
  read: async () => ({ ready: this.index.ready, pending: this.queue.length, lastRunAt: this.lastRunAt }),
});
// Now served as state["<your plugin id>"] on every authenticated GET /

read is called on every authenticated GET /, concurrently with every other extension's, and raced against the read budget from the plugin's advanced settings. A read that overruns, rejects, or resolves to anything but a JSON object is served as null for your extension and logged to the console, once, until a read succeeds again; nothing else in the response is affected. Keep it cheap: compute in the background and hand back the latest result. The schema, if you give one, is merged into the GET / response schema at /openapi.yaml, so clients can rely on the shape; without it the namespace is documented as a free-form object. Either way the published entry also admits null, since that is what a failed read is served as. unregister() removes both.

Known extensions

• Periodic Notes: Adds support for periodic notes

Contributing

See CONTRIBUTING.md. If you want to add functionality without modifying core, consider building an API extension instead — extensions can be developed and released independently.

Credits

Inspired by Vinzent03's advanced-uri plugin, with the goal of expanding automation options beyond the constraints of custom URL schemes.