Repository navigation
Releases
v0.2.1 — consent page redirect fix
Fixes sign-in from claude.ai and ChatGPT in Chrome. If you run --public-url, upgrade from 0.2.0: pipx upgrade nodus-mcp-server.
Fixed
-
Allow on the consent page did nothing. The page's
Content-Security-Policyhadform-action 'self'. Chrome appliesform-actionto the redirect that follows a form submit, so the 302 back to the app's callback (e.g.https://claude.ai/api/mcp/auth_callback) was blocked without any message. This is what happened:- The grant was made, but the app never got its code.
- A second click found the request already used.
The consent page now allows
'self'plus the origin of the app it is approving, and nothing else. An origin containing CSP punctuation falls back to'self'.
stdio, and --http without --public-url, are unaffected.
Full details: CHANGELOG.md · PyPI: https://pypi.org/project/nodus-mcp-server/0.2.1/
v0.2.0 — sign-in for --http
Sign-in for --http, so the server can sit behind a public URL that claude.ai and ChatGPT reach. Off unless you turn it on. stdio, and --http without --public-url, behave exactly as in 0.1.16.
pipx upgrade nodus-mcp-server
nodus-mcp-server set-passphrase
nodus-mcp-server --http --port 8765 --public-url https://<your-domain>Then add https://<your-domain>/mcp as a connector in claude.ai (choose DCR) or ChatGPT (OAuth). Your browser opens a consent page on your own server. Pick what the app may do and enter your passphrase.
Added
- With
--public-url(orNODUS_MCP_PUBLIC_URL), the server is its own OAuth 2.1 authorization server, built on themcpSDK's handlers. It provides:- dynamic client registration: open, capped at 50 clients, rate-limited;
- authorization code with PKCE (
S256only) and a consent page; - token, refresh and revocation endpoints, plus both metadata documents.
/mcpthen needs a bearer token.- Scopes per tool:
memory.read(the only one ticked by default),memory.write,workflow.run,exec. An app sees and can call only the tools its token allows. nodus_execis not offered over the public URL unless you pass--allow-remote-exec.- Tokens:
- Opaque, and stored only as hashes.
- Access tokens last 1 hour. Refresh tokens last 30 days and rotate on every use, and replaying an old one signs the app out.
- A new passphrase signs every app out.
nodus-mcp-server grants list | revoke <id>shows and ends connected apps.- With
--public-url, the MCP transport checksHostandOrigin.
Changed
- The
--httphelp text says streamable HTTP at/mcp, not SSE. - The startup message no longer tells you to point ChatGPT at a localhost URL.
Full details: CHANGELOG.md · PyPI: https://pypi.org/project/nodus-mcp-server/0.2.0/
v0.1.16
Dependency-floor repair plus the repo's first CI.
Changed
mcpfloor raised to>=1.22.0(was>=1.8). mcp below 1.22 importseval_type_backportfrompydantic._internal._typing_extra, which current pydantic (2.14) no longer exposes, soimport mcpfailed at collection — the declared floor did not install. 1.22.0 is the lowest release that imports with current pydantic and still carriesStreamableHTTPSessionManager. The suite passes against it.nodus-langfloor raised to>=5.15.1(was>=5.15.0). 5.15.0's bareimport tomllibbreaks every Python 3.10 install (nodus-lang #893); 5.15.1 is the first release that imports on 3.10, so the floor must be 5.15.1 for this package'srequires-python = ">=3.10"to hold.
Added
- A CI workflow (#9): a matrix of OS {ubuntu, windows} × Python {3.10, 3.13} × mcp {1.x, 2.x} plus the declared floor, and a package job that builds the wheel and completes a real
initializehandshake against the installed server. - A
CLAUDE.mdorienting the repo within the Nodus ecosystem.
These floors were found by adding CI and re-running it against nodus-lang 5.15.1: the general 3.10 cells went green (confirming nodus-lang #893), which unmasked the mcp/pydantic break that the tomllib failure had been hiding.
v0.1.15
Security
- This server has no authentication, and the README now says so. Anyone who can
reach--httpcan use every tool: read, change and delete memory, and run code in
thenodus_execsandbox. The ChatGPT setup told readers to expose it through ngrok
without saying that. If you have a tunnel running, stop it when you are not using it. nodus_execcould read and change the server's environment variables when
installed with nodus-lang below 5.0.0, whichnodus-lang>=4.0.5allowed and which
an older install keeps through an upgrade.allow_envwas never set, so it took
that version's default of on. It is now denied explicitly and the floor is raised.
Files, network and subprocess were already denied.nodus_resume_workflowtook any string as agraph_id, and nodus-lang builds
a file path from it, so../../xloaded a.jsonfile outside the state directory.
The contents were not returned, but the reply showed whether the file existed. An
id must now be letters, digits,_or-.
Fixed
nodus_resume_workflowfailed on the built-in workflow withUndefined variable: topic, from every checkpoint. Params were injected as globals, and a resume reloads
the workflow with none.researchnow declarestopicas a workflow parameter,
which nodus-lang stores with the run and restores on resume.serverInfo.versionreported the mcp SDK's version, not this package's.- The wheel no longer installs a top-level
testspackage intosite-packages.
Changed
- Workflow params must now be declared.
nodus_run_workflowbinds them to the
workflow's parameters, so a missing or unknown param is an error, and a custom
workflow must be writtenworkflow name(param) { ... }. Goals are unchanged.
Runs saved by 0.1.14 or earlier cannot be resumed; they never could. - Requires
nodus-lang>=5.15.0, which keeps a run's bounds across a resume. - The version now has one source,
nodus_mcp_server.__version__.
v0.1.14
Fixed
-
Works with the mcp SDK 2.x (#3). The server registered its two handlers
through the 1.x decorator API (@app.list_tools()/@app.call_tool()),
which mcp 2.0 removed, so with mcp ≥ 2.0 installedimport nodus_mcp_server.serverfailed — both transports. 0.1.13 pinnedmcp<2as
the stopgap; the pin is lifted tomcp>=1.8.One implementation with a branch per SDK major, chosen once at import
(_SDK_V2): on 2.x the handlers go throughadd_request_handler, take
(ctx, params)and return result models; 1.x is unchanged. The same port as
nodus-mcp 0.1.4 (nodus-mcp#11), so the two servers do not drift. Both
branches are driven bytests/test_http_transport.py's realinitialize→
tools/listthrough the ASGI app: 27/27 under mcp 1.27 and under 2.2.
v0.1.13
Fixed
--httpmode answered every request with a 500. 0.1.11 mountedStreamableHTTPSessionManagerand never entered itsrun(), so every POST to/mcpraisedRuntimeError: Task group is not initialized. The process started and printed its URL, which is all anyone checked. The manager now runs for the app's lifespan.tests/test_http_transport.pycompletes a realinitialize→tools/listthrough the app and pins the unrun shape as a 500. stdio mode was never affected.
Changed
mcppinned>=1.8,<2. mcp 2.0 removed the decorator API this server uses; a fresh install resolved 2.2.0 and failed at import in both modes. The 2.x port is #3.
Verified against the published wheel: fresh venv, initialize → tools/list (7 tools) → nodus_exec("print(6i * 7i)") → 42, over HTTP.
v0.1.9 — HTTPS/SSL support (no ngrok)
What's new
HTTPS / SSL support — no more ngrok
Pass --ssl-cert and --ssl-key to run over HTTPS locally. Works with mkcert locally-trusted certificates.
One-time setup:
winget install FiloSottile.mkcert
mkcert -install
mkcert 127.0.0.1 localhost
Start the server:
nodus-mcp-server --http --port 8765 --ssl-cert 127.0.0.1+1.pem --ssl-key 127.0.0.1+1-key.pem
The URL printed will be https://127.0.0.1:8765/sse — point ChatGPT Desktop at that directly. No tunnel needed.
Changes
- Added
--ssl-certflag (path to certificate file) - Added
--ssl-keyflag (path to private key file) - URL output switches to
https://automatically when SSL is configured - All other behaviour unchanged
Install / upgrade
pipx install nodus-mcp-server --force
v0.1.8 — HTTP/SSE transport for ChatGPT Desktop
What's new
HTTP/SSE transport mode
nodus-mcp-server now supports both MCP transport modes:
--stdio— Claude Desktop (spawned subprocess, unchanged)--http --port 8765— ChatGPT Desktop and any HTTP/SSE MCP client
Start the HTTP server:
nodus-mcp-server --http --port 8765
Then point ChatGPT Desktop at: http://127.0.0.1:8765/sse
Memory is shared between both modes — same SQLite database at ~/.nodus-mcp-server/data/memory.db.
Changes
- Added
--httpflag (Starlette + uvicorn +mcp.server.sse.SseServerTransport) - Added
--hostflag (default127.0.0.1) - Default port changed from 8080 to 8765
- README updated with full ChatGPT Desktop setup instructions
- Requires nodus-lang >= 4.0.5
Install / upgrade
pipx install nodus-mcp-server --force