What are your favourite MCP servers?

[markmcd]

What MCP servers are you using with the Gemini CLI?

Have you found anything surprisingly effective, or cool?

Tell us all about them!

[markmcd]

I'll go first - I used Gemini CLI to make a (really basic) MCP for search arXiv: https://github.com/markmcd/arxiv-search-mcp

The arXiv servers don't allow much traffic but it was fun to use with Chirp to generate an audio summary of research papers.

[centminmod]

Cloudflare ones https://developers.cloudflare.com/agents/model-context-protocol/mcp-servers-for-cloudflare/

Like

• https://github.com/cloudflare/mcp-server-cloudflare/tree/main/apps/docs-vectorize
• https://github.com/cloudflare/mcp-server-cloudflare/tree/main/apps/autorag

I also made Gemini CLI itself a MCP server via wrapper so I can add it to Claude Code so Claude Sonnet 4 and Gemini 2.5 models can be friends https://www.threads.com/@george_sl_liu/post/DLa8ZFNTYlV?xmt=AQF0_qghMaTOg0dW6z-UQ0899VkHOIaIFIC3P9Ci76ElMA 😎

[markmcd]

Those CloudFlare services look handy.

Can you share anything about how you set up the CLI-as-MCP?

I've been wondering if I can use the CLI as an MCP service in order to have local tools that represent other projects - e.g. if working on a UI app, but I need info about a downstream API, have that set up as a separate CLI MCP and query from the UI project. Might be overkill but it's something I want to try at some point.

[centminmod]

I fed Claude Code MCP docs and asked it to create a Gemini CLI MCP wrapper and it did it's thing 🤓 It created the mcp_server.py MCP file and gave me instructions how to install it.

claude mcp add gemini-cli /path/to/gemini-cli-mcp/.venv/bin/python /path/to/gemini-cli-mcp/mcp_server.py

Since I have added Gemini code, planning and solution review MCP tools (custom Gemini CLI -p prompts) that can also be called https://www.threads.com/@george_sl_liu/post/DLglV6Khe2d

[Manamama]

I am testing the "sequential thinking" one. Supposedly it makes it structure its thoughts better, but I did not find it to be the case that much, when I tested Gemini against my somewhat counter-factual puzzles, e.g. this one :
River puzzle🚣‍♂️, goat 🐐, cabbage 🥬, wolf 🐺: Here is a logic puzzle. Human needs to carry a cabbage, a goat, and a wolf across a river. Human can only carry one item at a time with human in the boat. Human cannot leave the goat alone with the cabbage, and human cannot leave the cabbage alone with the wolf. The human prevents danger when present, so any pair can be left safely as long as the human is with them. How can human get everything the other side?

In Gemini's own words how that tool works:

Here's a breakdown of its key features and parameters:

Key Features:

• Flexible Thinking: You can adjust the estimated total number of thoughts, question or revise previous thoughts, and add more thoughts even after reaching what seemed like the end.
• Adaptability: It allows for expressing uncertainty, exploring alternative approaches, and branching or backtracking in the thought process.
• Hypothesis-Driven: It supports generating and verifying solution hypotheses based on the chain of thought steps.

Parameters:

• thought: Your current thinking step. This can be an analytical step, a revision, a question, a realization, a change in approach, a hypothesis, or a verification.
• nextThoughtNeeded: A boolean indicating whether more thinking steps are required.
• thoughtNumber: The current thought's sequential number.
• totalThoughts: Your current estimate of the total thoughts needed for the task. This can be adjusted.
• isRevision: A boolean indicating if the current thought revises a previous one.
• revisesThought: If isRevision is true, this specifies which thought number is being reconsidered.
• branchFromThought: If branching, this indicates the thought number from which the current branch originates.
• branchId: An identifier for the current branch, if any.
• needsMoreThoughts: A boolean to indicate if more thoughts are needed even if the estimated totalThoughts has been reached.

In essence, it's a tool to help structure and manage a multi-step problem-solving process, allowing for iterative refinement and adaptation.

Result:
My rebuke was needed, because Gemini still hallucinates that the bad wolf must eat the goat, etc. into fake constraints, so of course it then missolves it, as most AIs do anyway, without any special tricks. But then indeed, after lots of thinking (61 steps seemingly):

1. Human takes the Cabbage across.
   * Bank A: {Goat, Wolf}
   * Bank B: {Human, Cabbage}
   * Reasoning: Goat and Wolf are left alone on Bank A, which is safe according to the given constraints (they don't eat each other).

2. Human returns alone to Bank A.

   • Bank A: {Human, Goat, Wolf}
   • Bank B: {Cabbage}

3. Human takes the Goat across.

   • Bank A: {Wolf}
   • Bank B: {Human, Cabbage, Goat}
   • Reasoning: Wolf is left alone on Bank A, which is safe. Human is with Cabbage and Goat on Bank B, preventing danger.

4. Human takes the Cabbage back to Bank A.

   • Bank A: {Human, Cabbage}
   • Bank B: {Goat, Wolf}
   • Reasoning: Goat and Wolf are left alone on Bank B, which is safe according to the given constraints.

5. Human takes the Wolf across.

   • Bank A: {Cabbage}
   • Bank B: {Human, Goat, Wolf}
   • Reasoning: Cabbage is left alone on Bank A, which is safe. Human is with Goat and Wolf on Bank B, preventing danger.

6. Human returns alone to Bank A.

   • Bank A: {Human, Cabbage}
   • Bank B: {Goat, Wolf}

7. Human takes the Cabbage across.

   • Bank A: {}
   • Bank B: {Human, Goat, Cabbage, Wolf}

Maybe I am still doing it the wrong way?

Ver. 1.2

[markmcd]

Haven't tried it yet but this one looks pretty: https://21st.dev/magic

If nothing else it's nice to have an alternative to the style of UI that Gemini generates by default.

[Manamama]

Puppeteer MCP works. I needed to add these interim tips though for Gemini: https://gist.github.com/Manamama/8804d0a03d6a6782a350fb986caaf320