Joplin MCP for Claude & ChatGPT

io.github.kogelerPublished by GitHub account kogelerv1.8.1Updated Nov 15, 2026

Self-hosted Joplin MCP server with built-in OAuth for hosted Claude and ChatGPT Actions

The publisher's own server: Claude Platform Docs

Overview

AI-generated overview

Self-hosted Joplin MCP server that lets assistants search, read, create, update, tag and trash notes in your own Joplin knowledge base.

What you can do
  • Search, read, create, update, move, tag and trash notes and notebooks
  • Upload, download and replace attachments and traverse their relationships
  • Serve typed MCP tools over local stdio or Streamable HTTP
  • Expose authenticated ChatGPT Actions and built-in GitHub OAuth for hosted Claude
  • Pull notebooks to Markdown files, inspect a three-way diff and dry-run, then push back
  • Deploy headless Joplin Terminal with recurrent sync as rootless systemd services
When to use it
Use it when the assistant runs elsewhere, such as hosted Claude or ChatGPT, when Joplin runs headless on a server, or when you need attachment handling and reviewable Markdown changes with diffs and Git history.
Requirements
A local Joplin instance with the Web Clipper/Data API enabled (default port 41184), or a headless Joplin Terminal deployment. A Joplin Web Clipper token supplied via JOPLIN_TOKEN or a protected token file. Python 3.13+ or a native executable; the headless installer needs Linux with systemd user services plus Node.js/npm. Hosted access needs a public HTTPS hostname, and Claude OAuth needs a GitHub OAuth App with Client ID, Client Secret and numeric user ID.
Before you install
The JOPLIN_TOKEN grants access to your notes; keep it in a protected file or environment variable, since the legacy --token argument exposes it in the process list and IDE configuration. Create, update, move, tag and delete operations write to your notes; resource and tag deletion is destructive because Joplin has no trash endpoint for them. Exposing the adapter publicly requires TLS and separate bearer tokens, and the Joplin Data API should stay on loopback. Hosted Claude and ChatGPT routes…

Installation

In SourceWeft

  1. Open Joplin MCP for Claude & ChatGPT in the dashboard and add it to a workspace.
  2. Enable the server for the chats that should use its tools.

Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.

Other MCP clients

Follow the launch instructions in the repository.

README

Joplin MCP for Claude and ChatGPT

joplin-md-sync gives AI assistants controlled access to your own Joplin knowledge base. Search, read, create, update, move, tag, and trash notes from hosted Claude with a GitHub login, a private Custom GPT, or any local stdio or Streamable HTTP MCP client without moving your notes into another hosted note service.

  • Built-in OAuth for hosted Claude. Claude on claude.ai connects straight to Joplin on your own server through your GitHub login. OAuth is part of the server itself, so there is no separate OAuth proxy or gateway to deploy and configure, and the Joplin Data API stays on loopback.
  • ChatGPT today. A private Custom GPT works through authenticated Actions. OpenAI retires Custom GPTs on 2026-12-11; a ChatGPT MCP connector through the same built-in OAuth is planned as the replacement.
  • One installer for the whole stack. A single command sets up headless Joplin Terminal, recurrent sync, MCP, Actions, and Claude OAuth as rootless systemd services. It asks for the values it needs, hides secrets, generates keys and tokens, and verifies the result.

Run it beside Joplin Desktop, or deploy the complete headless stack on your own Linux host. A separate Markdown sync workflow is available when a coding agent needs reviewable files, three-way diffs, and Git history.

Website · Connect Claude · Connect ChatGPT · MCP: local stdio & HTTP · Deploy headless

What you get

  • Claude with GitHub sign-in. The built-in /oauth/mcp interface lets hosted Claude, currently its only supported OAuth client, sign in with one configured numeric GitHub account on the existing service port. The installer configures it from a few answers; CLI flags and private secret-file options do the same by hand.
  • Joplin inside ChatGPT. A generated OpenAPI contract exposes authenticated Actions for current notes, notebooks, tags, and search to a private Custom GPT.
  • Typed MCP tools. Local stdio and Streamable HTTP transports give compatible clients structured note operations plus attachment upload, download, replacement, and relationship traversal.
  • A headless Joplin service. The rootless Linux installer deploys Joplin Terminal, recurrent sync, and the combined MCP, OAuth, and Actions adapter as coordinated systemd user services. Joplin Desktop does not need to remain running.
  • Your storage and encryption choices. Use filesystem, Nextcloud, WebDAV, Dropbox, OneDrive, S3, Joplin Server, or Joplin Cloud, with existing Joplin end-to-end encryption where the selected topology supports it.
  • Reviewable Markdown when you need it. Pull notebooks to ordinary files, let an agent work with repository context, inspect a three-way diff and dry-run, then push the verified result back to Joplin.

Joplin remains the source of truth. The adapter uses its documented local Data API and never edits the Joplin database, profile, or sync target directly.

Joplin Desktop also ships its own MCP server for clients on the same computer. Use this project when the assistant runs elsewhere, such as hosted Claude or ChatGPT, when Joplin runs headless on a server, or when you need attachments and reviewable Markdown changes. See the comparison.

Fast path: headless Joplin for Claude and ChatGPT

This topology keeps Joplin and the adapter on your own Linux host. Only the authenticated OAuth or Actions routes need to reach the assistant over HTTPS; the Joplin Data API stays on loopback.

1. Install Joplin and the agent service

The host needs Linux with systemd user services, Python 3.13.5 or newer, and Node.js/npm. This interactive example connects the new headless profile to an existing Nextcloud Joplin sync target:

bash
set -o pipefailcurl --proto '=https' --tlsv1.2 --fail --silent --show-error --location \  'https://raw.githubusercontent.com/kogeler/joplin-md-sync/main/scripts/joplin_terminal_service/install_joplin_terminal.py' \  | python3 - \      --sync-target nextcloud \      --sync-location 'https://cloud.example.com/remote.php/dav/files/user/Joplin' \      --sync-username 'user'

The installer asks for secrets through hidden prompts, offers GitHub OAuth for Claude, handles existing Joplin E2EE keys, verifies release checksums, and creates:

text
joplin-terminal.service    Joplin profile, Data API, and recurrent syncjoplin-md-sync.service     MCP, Claude OAuth, and ChatGPT Actions on one guarded adapter

It also generates separate protected credentials for Joplin, MCP, and Actions, plus the OAuth signing key when OAuth is enabled. For ready-to-run Nextcloud, custom S3, and uninstall commands, see Joplin Headless Service. The guide also explains installed paths, every installer option, and upgrades.

2. Publish the narrow HTTPS routes

From a trusted HTTPS hostname, route only what your assistant needs to the adapter on 127.0.0.1:8765: the OAuth routes for Claude, /api/gpt/v1/* for ChatGPT, or both. Do not expose the upstream Joplin Data API. Keep /mcp private unless a remote MCP client needs it; public MCP requires its own bearer token and TLS.

The supported boundaries and deployment choices are documented in Self-hosted deployment.

3. Connect Claude

Register a GitHub OAuth App, then answer yes to the installer's Enable GitHub OAuth for Claude? question or rerun it with --github-oauth. It asks for the public HTTPS origin, the app's Client ID and Client Secret, and your numeric GitHub user ID, then enables OAuth on the running service and checks it. Add https://YOUR_HOST/oauth/mcp as a custom connector in Claude and sign in with GitHub. The exact routes, app settings, and checks are in the Claude connector guide.

4. Connect a private Custom GPT

From a checkout matching the deployed release, run the setup assistant:

bash
git clone --depth 1 --branch v1.8.1 \  https://github.com/kogeler/joplin-md-sync.gitcd joplin-md-syncpython3 scripts/prepare_chatgpt_action.py

Enter the public hostname and the generated Actions token when prompted. The assistant validates TLS, authentication, and live read operations, then writes the OpenAPI file for the Custom GPT editor. Continue with the exact GPT instructions and acceptance test in ChatGPT Actions setup.

Use Joplin from an MCP client

For a local Joplin Desktop instance, enable Tools > Options > Web Clipper and point an MCP-capable editor at the downloaded native executable:

json
{  "mcpServers": {    "joplin": {      "command": "/absolute/path/to/joplin-md-sync-linux-amd64",      "args": ["mcp", "stdio", "--token-file", "/home/USER/.config/joplin-md-sync/token"]    }  }}

Store the token once in a private file outside synchronized folders. The directory is 0700, the file 0600, and the token is read without echo or shell history:

bash
install -d -m 0700 ~/.config/joplin-md-sync(umask 077; read -rs -p 'Joplin token: ' token &&  printf '%s\n' "$token" > ~/.config/joplin-md-sync/token)

A launcher that exports environment variables can set JOPLIN_TOKEN instead and pass neither token option. Use exactly one source: a command-line option together with a non-empty JOPLIN_TOKEN is rejected. The older "--token", "<token>" arguments still work for existing configurations, but they place the token in the IDE configuration, in the process list, and in any agent command line that forwards the MCP configuration.

The local stdio process connects to Joplin on port 41184 by default; add "--port", "PORT" to the arguments when Joplin uses another local port. For a long-running or remote adapter, install the package and start Streamable HTTP instead. Stdio opens no MCP or Actions port and therefore needs no bearer token for those interfaces; its Joplin token authenticates only to Joplin:

bash
pipx install "joplin-md-sync==1.8.1"export JOPLIN_TOKEN=...joplin-md-sync mcp serve

Connect the HTTP client to http://127.0.0.1:8765/mcp with transport type streamable-http.

The service can start while Joplin is offline and recovers on later calls. Create operations reject an existing notebook, note, tag, or resource identity instead of creating accidental duplicates. Remote MCP deployment requires a separate protected bearer token and TLS. See the complete MCP API reference.

Use Joplin notes as reviewable files

Choose the Markdown workflow when an agent needs repository context, broad transformations, an exact diff before writes, or selected notebooks in Git:

bash
export JOPLIN_TOKEN=...joplin-md-sync init --root ./notes --mode remote-firstjoplin-md-sync pull --root ./notes --json
# Let an agent edit the managed Markdown files, then review the result.joplin-md-sync diff --root ./notes --three-way --unifiedjoplin-md-sync push --root ./notes --dry-run --jsonjoplin-md-sync push --root ./notes --json

Managed note files carry a one-line Joplin identity header. Base snapshots, conflicts, journals, backups, and downloaded resources stay under the ignored .joplin-sync/ directory. Start with the copyable agent notes repository template or the Markdown quick start.

Control and failure behavior

  • The Joplin Data API remains private on loopback in the headless topology.
  • Joplin, MCP, and Actions credentials are distinct and read from protected files or the environment. Local mcp stdio takes its Joplin token from a protected --token-file or JOPLIN_TOKEN; the compatibility --token argument exposes it in the process list.
  • Direct API writes are sent once. An ambiguous timeout is reported instead of being replayed and possibly duplicated.
  • Creating an occupied notebook, note, tag, or resource identity returns an explicit *_ALREADY_EXISTS error with the existing ID and recommended update tool.
  • Note and notebook deletion uses Joplin trash. Resource and tag deletion is explicitly marked destructive because Joplin has no trash endpoint for them.
  • Divergent file edits produce a conflict bundle instead of a silent overwrite.
  • File deletion propagation is off by default; interrupted mutations are journaled and recoverable; diff never mutates state.

The exact test-backed guarantees live in the contract catalog. Agents operating a Markdown workspace should also follow AGENTS.md.

Installation options

Python installations require CPython 3.13 or 3.14 on Windows or Linux:

bash
python -m pip install "joplin-md-sync==1.8.1"# or: pipx install "joplin-md-sync==1.8.1"

Native executables from GitHub Releases include Python and the standard service runtime. A Python zipapp is also available for a prepared Python environment:

PlatformAsset
Linux AMD64joplin-md-sync-linux-amd64
Linux ARM64joplin-md-sync-linux-arm64
Windows AMD64joplin-md-sync-windows-amd64.exe
Python zipappjoplin-md-sync.pyz

Download them with SHA256SUMS.txt from the latest release.

Interface guide

GoalInterface
Connect hosted Claude with GitHub sign-inBuilt-in OAuth MCP at /oauth/mcp
Work with current Joplin notes in ChatGPTAuthenticated ChatGPT Actions
Give an editor or assistant typed note toolsMCP
Run without a desktop sessionHeadless Joplin Terminal plus MCP, OAuth, or Actions
Review every broad agent edit before applying itMarkdown workspace
Keep selected notebooks in GitMarkdown workspace

Compare consistency models and deployment choices in Choose an agent interface.

Development

bash
make venvmake venv-devmake checkmake cimake test-livemake packagemake docs-auditmake docs-screenshotsmake help

The standard installation includes the HTTP and GitHub OAuth runtime; CLI flags enable the interfaces. Development, test, package, and docs tools use separate dependency groups from the single hash-verified uv.lock. See Development and Dependency maintenance. On Linux AMD64, make test-live provisions a checksum-pinned Joplin Desktop 3.6.15 binary and isolated temporary profile; it never uses the user's Joplin process or token.

Releases and license

Git tags use vX.Y.Z. Release assets include wheel, sdist, zipapp, native executables, and checksums; the same wheel and sdist are published to PyPI through Trusted Publishing. See the changelog.

MIT licensed. This is an independent project and is not affiliated with or endorsed by the Joplin project.

Source: README.md at commit 80d501e

Tools

0
SourceWeft started this package once in an isolated sandbox, without credentials, to list its tools. That failed: the package did not install.

Version history

1
  1. v1.8.1LatestNov 15, 2026

Score

50% of the applicable checks met

How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.

  • Reached in the last 30 days25 pts

    Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.

  • Auth declared10 pts

    Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.

  • Tools listed15 pts

    Answer tools/list without a credential, or ship a package that starts in a sandbox without one.

  • DescriptionRequired15 pts
  • Logo10 pts
  • Documentation15 pts
  • License10 pts

Other Claude integrations