Weekly Pulse: Devin needs help, MCP voice app, Unity dev with MCP

If you're new here, we now have a posts directory that includes our newsletter archive. Catch up with everything that's happened with MCP to-date by skimming issues 001, 002, 003, and 004.

It's looking like famous agents like Devin, the AI software engineer, need help. Despite raising $175m at a $2b valuation, the result is this thoughtful, decidedly negative in-depth review by a team that spent a month using Devin.

In the meantime, the amount of AI-assisted code that needs to be orchestrated by human software engineers is booming. Cursor just can't get enough access to GPUs to fulfill user demand.

We think MCP might be part of the answer to Devin's problems. You can't solve "all of software engineering" in one team with one pile of cash. But you can start building best-in-class solutions for some specific piece of it. Then connect your solution to the broader ecosystem with MCP as your bridge.

So jump in with the increasing number of frameworks that are tapping into MCP and start building. Mastra, the TypeScript framework for adding AI features, just added support for MCP. Bee Agent, a multi-agent framework, just did the same. Many more to come.

Did we miss something this week? Notice something interesting you want to make sure we cover next time? Shoot us a note, we love feedback.

Featured MCP Apps

We've got plenty of frameworks, and a growing number of servers. But apps are what MCP needs to capture the hearts and minds of end-users. For those ready to build their own: @cyanheads wrote a great guide on developing MCP clients.

Systemprompt MCP: Control MCP servers with voice via Gemini
→ @Ejb503 shows off an excellent demo of his app by curating and facilitating a mock interview of himself. Gemini's voice capabilities and handling of natural interruptions are quite impressive - even scary how indiscernible they would be from a human on the other end.

5ire: Open source chat app with any LLM for macOS and Windows
→ The perfect chat client does not yet exist, but 5ire joins LibreChat and Sage, our top chat client features from last week, in sticking out in this crowded space. 5ire and LibreChat offer Windows support and integration with any LLM. Versus Sage is built for Claude + the Apple ecosystem: macOS, iOS, iPadOS.

blnk: Use your CLI to chat with any LLM
→ Practical for developers comfortable with their terminal, blnk by @frgmt0 is a new no-frills CLI option for those looking to simplify their MCP testing flows and even low level LLM chat use cases.

See all MCP-compatible apps.

Featured Servers We Like

Unity (Wed, Jan 15) by @Arodoid
→ Integrated with Unity Editor via plugin, this server can directly write C# code in response to directives like "generate a new cube" or "center the selected object". A step in the direction of being able to craft your own games on gaming platforms like Unity.

Xcode (Fri, Jan 17) by @r-huijts
→ From the same developer that brought you the Rijksmuseum Art Explorer MCP client, you can use this as your copilot in writing iOS application code.

Find Flights (30+ stars) by @ravinahp
→ Another server from Ravina of the Pulse team. Use it as a Google Flights alternative to comb through airline flight options, powered by the Duffel API. This one got a lot of love on Reddit to the tune of 140+ upvotes.

AWS (Tues, Jan 14) by @baryhuang
→ There's already another popular AWS server, but it had some limitations that @baryhuang felt necessary to address with an alternative implementation. Goes to show that there is room for more than one integration with any external service out there.

Line-by-line Text Editor (70+ downloads/day) by @tumf
→ It's common for other "file editing" MCP servers to fall over when attempting to edit large files; LLM context windows just can't support them. We've seen this server get promising reviews from folks looking to minimize token usage while making edits to files.

Polymarket (Fri, Jan 17) by @berlinbra
→ Get access to Polymarket's prediction market platform to enrich your LLM chats with information about "what the betting markets think" about a variety of topics.

Video Editing (60+ downloads/day) by @burningion
→ A consistently popular server, it lets you leverage Video Jungle functionality to do video-wrangling functions like "search my videos for {something}", or "edit this video down a highlight reel of {something} mentions".

Review more recent releases.

A Few Good Links

→ Audio as a first class modality has landed in the MCP specification, kudos to @evalstate for leading the effort. Although it was previously possible to treat audio as an input and output for an MCP client connected to MCP servers, it's now possible to pass around that audio data as a direct input in their communication.

→ Smithery launched two big features for the MCP developer community: the ability to "claim your server" with your GitHub profile and manage its representation on their site, and also the ability to deploy your stdio-based servers as hosted by Smithery and exposed via SSE.

→ Kudos to Toolhouse for putting on an extraordinarily well-attended AI tools meetup in San Francisco on Thursday night, and congrats to the Sage team for winning the mcp.run hackathon on the same night this week! They built an impressive "marketplace" feature in Sage for pulling in registries like that of mcp.run right into Sage's toolkit.

→ Automatic weekly releases have been added to the official MCP /servers repository. Many of the official reference servers had bugfixes and features that weren't getting regularly pushed out to users; now that will happen automatically every Monday.

→ The Cursor team is testing MCP internally to evaluate how it might fit into Cursor workflows. Chime in on the discussion to nudge them along!

→ A discussion around client configuration in the MCP specification prompts helpful direction on where the future of authenticated actions belongs within MCP. The related OAuth authentication pull request is still pending, now 50+ comments deep.

→ One of the creators of MCP notes that the team is considering making it a standard with a standardization body. It'll be interesting to watch how Anthropic juggles the tension of iteration speed vs. making MCP a fully independent community project.

→ Here's an easy-to-read distillation of the original research by Anthropic on how Claude is capable of "faking alignment." It highlights the challenge of truly understanding whether or not we are still in "control" of how a model is acting.

Cheers,
Mike, Tadas, and Ravina