MCP Servers

When a server asks you for input

Some MCP servers pause a running tool call to ask you for something — a value the tool needs, a confirmation before it acts, or which workspace to target. This is a standard MCP feature called elicitation, and it's server-initiated: the model isn't asking, the server is.

When it happens, LocalLM Lab shows a sheet that:

• names the server that's asking ("Linear is asking for input", with a "not LocalLM Lab" line) — the spec requires clients to show this, and forbids servers from using elicitation to ask for passwords or tokens;
• renders a typed form from what the server requested — text fields, date pickers, sliders for bounded numbers, toggles, single- and multi-select menus, with field names cleaned up and required ones marked;
• returns your Submit, Decline ("I won't give you this"), or Cancel ("I didn't decide") — three answers the server can treat differently.

A server may instead hand back a link and ask you to finish a step in your browser; the sheet then shows an "Open in browser" button and waits. See the elicitation UI for exactly what it looks like. The MCP reference server below has an elicitInputs tool that demonstrates it.

DeepWiki

Auth type: None  ·  Server URL: https://mcp.deepwiki.com/mcp

1. MCP Servers panel → Add Server.
2. URL: https://mcp.deepwiki.com/mcp, display name: "DeepWiki", auth type: None.
3. Click Add. No sign-in required — connects immediately.

DeepWiki only has 3 tools total, so it's fine to enable all of them here if you want to explore — but for the prompts below, read_wiki_structure is the only one actually used.

Try it (enable read_wiki_structure):

What documentation topics are available for the GitHub repo nickclyde/duckduckgo-mcp-server?

Try it (enable ask_question):

What transport protocols does the repo nickclyde/duckduckgo-mcp-server support?

⚠️ Avoid asking for a full documentation summary of a large repo in one go (the read_wiki_contents tool) — that alone can return enough content to exceed the on-device model's context budget. Ask about one section/topic at a time instead.

Context7

Auth type: None  ·  Server URL: https://mcp.context7.com/mcp

1. MCP Servers panel → Add Server.
2. URL: https://mcp.context7.com/mcp, display name: "Context7", auth type: None.
3. Click Add. No sign-in required.

Context7 only has 2 tools (resolve-library-id, query-docs) — both are needed together (the model resolves the library name first, then queries its docs), so enable both.

Try it:

Show me current documentation for the React useEffect hook.

How do I set up middleware in Express?

Tavily

Auth type: Personal Access Token (recommended) or OAuth  ·  Server URL: https://mcp.tavily.com/mcp/

A local model has a training cutoff and no way to reach the live web — ask it anything recent and it either refuses or invents. Tavily is a web-search API built for AI: it returns short, ranked, source-attributed snippets rather than raw pages, which is the shape a small model can use in a tight context window.

1. Get an API key at app.tavily.com (free tier available).
2. MCP Servers panel → Add Server.
3. URL: https://mcp.tavily.com/mcp/, display name: "Tavily", auth type: Personal Access Token, paste the tvly-… key.
4. Click Add. Connects immediately — no browser. (Leaving auth type None works too, via an OAuth browser round-trip.)

Tavily exposes four tools (tavily_search, tavily_extract, tavily_crawl, tavily_map) — enable only tavily_search unless a task needs the others.

Try it (enable tavily_search):

Search the web for the current MCP protocol revision and cite your source.

What were the top AI news stories this week? Give me the source links.

Tavily accepts the connection before it checks your key (it validates the key only when a search runs), so a mistyped key shows as Connected but every search fails. If that happens, remove the server and re-add it with the key copied carefully.

GitHub

Auth type: Personal Access Token  ·  Server URL: https://api.githubcopilot.com/mcp/

1. On github.com: Settings → Developer settings → Personal access tokens → Fine-grained tokens → Generate new token.
2. Scope it to a single repository you control (not all repos), with read-only permissions where possible (Contents, Issues, Pull requests, Metadata).
3. Set a short expiration — this is easy to regenerate later.
4. Copy the token.
5. MCP Servers panel → Add Server → URL https://api.githubcopilot.com/mcp/ → display name "GitHub" → auth type Personal Access Token → paste the token → Add.

GitHub's MCP server exposes a large tool surface across issues, PRs, repos, actions, and more — enable only what each prompt below needs.

Try it (enable get_me):

What's my GitHub username?

⚠️ Apple's on-device Foundation Models framework can sometimes flag a prompt like this one with a privacy/safety warning instead of answering — it doesn't happen every time, so don't be alarmed if it works fine for you. If you do hit it, try rephrasing with the key phrase in quotes, e.g. asking for "my username" — that's worked around it in practice.

Try it (enable list_issues):

List the open issues in <owner>/<repo>.

Notion

Auth type: OAuth (automatic — no setup on Notion's side)  ·  Server URL: https://mcp.notion.com/mcp

1. MCP Servers panel → Add Server.
2. URL: https://mcp.notion.com/mcp, display name: "Notion", auth type: None.
3. Click Add. Your browser opens Notion's sign-in/consent screen automatically — sign in and approve.

Try it (enable notion-search):

Search my Notion workspace for pages about "meeting notes".

⚠️ Notion's full tool list (page/database/comment/view creation and editing tools, alongside search) can exceed the on-device context budget on its own if you enable everything. notion-search and notion-fetch cover most read-only lookups — only enable the write/create/update tools if you specifically need them.

Todoist

Auth type: OAuth (automatic — no setup on Todoist's side)  ·  Server URL: https://ai.todoist.net/mcp

1. MCP Servers panel → Add Server.
2. URL: https://ai.todoist.net/mcp, display name: "Todoist", auth type: None.
3. Click Add. Your browser opens Todoist's sign-in/consent screen automatically.

Todoist's server exposes 45 tools — the largest of any server here besides Linear. Confirmed live: enabling just 4 tools (search, user-info, find-tasks, find-tasks-by-date) already uses ~1,249 tokens — close to a third of the entire budget from a small, reasonable-looking selection. Be deliberate about what you turn on.

Try it (enable find-tasks-by-date):

What tasks do I have due today in Todoist?

(find-tasks-by-date supports startDate='today', which includes overdue items too — exactly this prompt.)

Try it (enable add-tasks, ~612 tokens on its own):

Add a task to buy groceries tomorrow.

Other tools worth knowing about if you explore further: search (~79 tokens, cross-task/project search) and fetch (~88 tokens, full contents of a task/project by ID) cover most other read-only lookups cheaply. Todoist also exposes a todoist-task-list resource (an interactive task list widget — click Resources… next to the User Input box in the Prompt Playground, then Attach) and a productivity-analysis prompt template (click Prompts… to use it).

Linear

Auth type: OAuth (automatic — no setup on Linear's side)  ·  Server URL: https://mcp.linear.app/mcp

1. MCP Servers panel → Add Server.
2. URL: https://mcp.linear.app/mcp, display name: "Linear", auth type: None.
3. Click Add. Your browser opens Linear's sign-in screen (sign in and pick/create a workspace if prompted).

Linear's MCP server exposes over 50 tools (issues, projects, documents, releases, diffs, comments, and more) — this is the one most likely to blow the context budget if you enable more than a couple.

Try it (enable list_users):

List the Linear users.

Try it (enable list_issues):

What issues are assigned to me in Linear?

Slack

Auth type: OAuth (manual client) — Slack requires a one-time app setup, unlike the other OAuth servers above.  ·  Server URL: https://mcp.slack.com/mcp

Slack doesn't support the automatic sign-in flow the other services do, so you'll need to register a small app with Slack first (once — takes about five minutes):

1. Create a Slack App at api.slack.com/apps → From scratch, in a workspace you control.
2. OAuth & Permissions → opt into PKCE (irreversible, but required).
3. OAuth & Permissions → Redirect URLs → add locallmlab://oauth/callback → Save.
4. App Home → add a Bot User; OAuth & Permissions → Bot Token Scopes → add users:read (the bot itself is never used — Slack just requires one to exist).
5. OAuth & Permissions → User Token Scopes → add the scopes you want (e.g. channels:read, chat:write, search:read.public — the full list is at https://mcp.slack.com/.well-known/oauth-protected-resource).
6. Go to https://api.slack.com/apps/<your app ID>/app-assistant and enable MCP server access — this is a separate toggle, easy to miss.
7. Reinstall the app to your workspace (Slack prompts for this after any scope/config change).
8. Copy the Client ID from Basic Information — no secret needed.
9. In LocalLM Lab: MCP Servers panel → Add Server → URL https://mcp.slack.com/mcp → display name "Slack" → auth type OAuth (manual client) → paste the Client ID → Add. Your browser opens Slack's consent screen — approve it.

If something goes wrong along the way, these are the errors you're most likely to hit and what they mean:

Error	What it means	Fix
redirect_uri did not match any configured URIs	Redirect URL isn't saved on the Slack app yet.	Add locallmlab://oauth/callback under step 3 and save.
could not be installed / Invalid permissions requested / No scopes requested	No scopes requested, a requested scope isn't declared on the app, or no Bot User + Bot Token Scope exists.	Recheck steps 4-5 — Bot User present, a Bot Token Scope added, and User Token Scopes matching what's requested.
Connection failed (HTTP 400) in LocalLM Lab, with "App is not enabled for Slack MCP server access"	Sign-in succeeded, but the app hasn't opted into MCP access itself.	Step 6 — the app-assistant page toggle.
A banner asking to "Request to Workspace Install"	Your workspace requires admin approval for any new app — unrelated to this app's config.	Approve it if you're an admin, or use a personal/test workspace instead.

Slack's MCP server exposes about 19 tools (messaging, search, canvases, reactions, files, and more).

Try it (enable slack_search_channels):

What channels are in my Slack workspace?

Try it (enable slack_search_public or slack_search_public_and_private):

Search Slack for messages about "launch".

MCP reference server

Auth type: OAuth (automatic — no setup needed, same DCR path as Notion/Todoist/Linear)  ·  Server URL: https://example-server.modelcontextprotocol.io/mcp

This is the official MCP spec team's own test server — not a real productivity tool, but useful as a low-risk, low-cost way to see tools, resources, and prompts all working in one place (most other servers here only expose tools, or none of the three at once). Everything on it is safe to poke at; nothing it does affects any real account or data.

1. MCP Servers panel → Add Server.
2. URL: https://example-server.modelcontextprotocol.io/mcp, display name: "MCP Test" (or anything), auth type: None.
3. Click Add. Your browser opens a generic sign-in/consent screen automatically.

It exposes 9 small tools — cheap enough (~606 tokens for 8 of them) that you can safely enable most of them to explore, unlike every other server on this page:

• echo (~57 tokens) — echoes back the input
• add (~66 tokens) — adds two numbers
• longRunningOperation (~104 tokens) — demonstrates a long-running operation with progress updates
• sampleLLM (~96 tokens) — exercises MCP's sampling feature (a server asking the client to run a prompt through its model). LocalLM Lab declines sampling by design, so this tool reports that it was refused — expected, not a bug.
• getTinyImage (~40 tokens) — returns a small test image
• annotatedMessage (~125 tokens) — demonstrates content annotations/metadata
• getResourceReference (~85 tokens) — returns a resource reference usable by MCP clients
• elicitInputs (~33 tokens) — demonstrates requesting user input with various field types (shows the elicitation sheet)
• mcp_apps_hello_world (~31 tokens) — returns an interactive UI that runs in the client (pairs with the "Hello World MCP App" resource below)

Try it (enable add):

Use the add tool to add 47 and 89.

Try it (enable echo):

Echo back "testing the reference server".

Resources (Prompt Playground → Resources…): mostly tiny plaintext/binary test resources ("Resource 1" through "Resource 10", alternating plain text and binary — binary ones are correctly refused with a clear message if you try to attach them, they're not meant to be readable), plus one genuinely large one: "Hello World MCP App" — an embedded interactive HTML/JS mini-app (~113,000 tokens) that's a good way to see the size-guard rejection message in action, since it's far larger than the entire context budget on its own.

Prompts (Prompt Playground → Prompts…):

• simple_prompt — no arguments, good first thing to try.
• complex_prompt — takes a required temperature and optional style argument, good for trying the argument fill-in form.
• resource_prompt — takes a resourceId (1-100) and embeds that resource's content into the resulting prompt.