beckon
Beckon blog

Open Code Review, FastMCP, ffmpeg-skill: 5 Repos for Coding Agents

ยท Beau Johnson

In shortFive trending GitHub repos that each give a coding agent a new job: FastMCP (tool maker), Stitch Skills (designer), last30days (researcher), ffmpeg-skill (video editor) and Alibaba's Open Code Review (code reviewer). I'd install Open Code Review first, in delegation mode, since it needs no new key.

In this video I count down five repos trending on GitHub this week that each hand your coding agent a new job: tool maker, designer, researcher, video editor and code reviewer. I ran every one I could with no API keys, and I say where I couldn't. It's for anyone using Claude Code, Codex or Cursor who wants to know which of these is worth installing.

What I cover in this video

  • 0:00 The five jobs, and the Alibaba result that opens the video.
  • 0:45 How I picked the repos and why the ranking works the way it does.
  • 1:32 No. 5, FastMCP: what one decorator does, and the 23-line server I wrote.
  • 3:02 No. 4, Stitch Skills: Google's design tool wired into your agent.
  • 4:33 No. 3, last30days: one command, one cited research brief.
  • 6:52 No. 2, ffmpeg-skill: why agents guess at FFmpeg and how this stops it.
  • 8:26 No. 1, Open Code Review: Alibaba's benchmark and how the pipeline works.
  • 11:30 The one I'd install tonight.

I pulled the daily and weekly GitHub trending charts on Sunday, Nov 15, kept only repos that give an agent a new job, and skipped anything from earlier roundups. Four are ranked by stars gained this week. FastMCP only shows up on the daily chart, so it starts at number five.

Tool maker and designer: FastMCP and Stitch Skills

FastMCP (from Prefect, 82 stars that day, over 28,000 total) turns a normal Python function into an MCP tool with one decorator. MCP is how your agent talks to tools, and writing a server by hand is a lot of boilerplate. The decorator reads your function: the name becomes the tool name, the type hints become the input schema, and the docstring becomes the description your agent reads to decide when to use it. Every call is checked before your code runs, so a word sent where you wanted a number gets bounced.

I wrote a 23-line server with one tool that counts TODO and FIXME comments in a folder. FastMCP's inspect command printed the schema it built, including defaults and which input is required. I didn't write any of that schema. Pointed at the official MCP Python SDK, it found 31 notes across 18 files. Prefect says FastMCP 1.0 was folded into that SDK in 2024 and that some version of it runs 70% of MCP servers (their number). PyPI stats showed about 1.7 million downloads in a single day.

Stitch Skills (Google Labs, 442 stars this week) is 16 skills in three plugins that connect your agent to Stitch, Google's AI design tool, through the Stitch MCP server. It works in Claude Code, Codex, Cursor, Gemini CLI and Antigravity.

  • Design makes screens, edits them, makes variants and can pull a design system out of existing code.
  • Build turns screens into React components, React Native or a Vite dashboard, and helps with shadcn.
  • Utilities cleans up your prompt and runs a build loop.

The React skill gates every step: download each screen's HTML and screenshot, look at the screenshot, pull the colors, fonts and spacing into a style guide, then build. The build loop uses a baton file: each run builds one page, writes the prompt for the next and hands it off. The caveats: it needs the Stitch MCP server and your own Stitch API key, which I didn't set up, so I only read the skill files. The last commit was in August, and Google says it is not an officially supported product.

Researcher and video editor: last30days and ffmpeg-skill

last30days (Matt Van Horn, 794 stars this week, almost 68,000 total) researches a topic across Reddit, X, YouTube, Hacker News, Polymarket and the web. You type /last30days and a topic. Your agent works out who matters (X handles, subreddits, the GitHub repo), searches every source at once, scores results by real engagement, merges the same story from different sites and writes one cited brief. Your session then knows all of it, so you can keep asking questions.

I ran it on "Claude Code Skills" with no keys. The research part finished in 4.9 seconds: 15 Hacker News stories, 12 Polymarket markets and 3 GitHub results. Reddit was blocked from my network, and the brief said so instead of pretending. Most of what came back was general Claude Code news, because the Hacker News search dropped the word "skills". It's fast and honest about what it missed, but the keyless version is thinner than the README makes it sound.

ffmpeg-skill (Shogo Kajiyama, 1,110 stars this week) is 42 FFmpeg tools for Claude Code, Cursor and Codex, plus an MCP server. No cloud, no API keys, just FFmpeg and Python on your machine. It exists because agents guess: the frame rate, a codec the file can't hold, then "done" without opening the result. A bad FFmpeg command often still finishes with no error.

Every job starts with probe, which measures the real file. The edit runs as a typed tool, never a raw shell command. Then it measures its own output, and if the picture changed, it makes a contact sheet the agent has to look at. The tool schema is generated from each script's own arguments, and the build fails if they drift. Over MCP it lists only the core 12 tools to save context.

On a 45-second test clip, I cut two 10-second pieces with no re-encode, fit them to 15 seconds vertical, added captions and took loudness from minus 20.7 to minus 14 LUFS. The check tool ran 12 checks for Shorts with 0 failed, and the final file was 15 seconds, 1080 by 1920, 30 frames a second.

Code reviewer: Alibaba's Open Code Review

Open Code Review (2,238 stars this week, over 52,000 total, Apache 2.0) is the reviewer Alibaba says its developers used for two years. On Alibaba's own benchmark, AACR-Bench, the same Claude Opus 4.6 model reviewed 200 real pull requests two ways:

  • Inside Claude Code: 5,980 comments, about 7% real issues, about 5.7M tokens and 13 minutes per review.
  • Inside Open Code Review: 889 comments, 34% real issues, about 385K tokens and a minute and a half per review.

Claude Code still found more real issues, 435 versus 301. That's the trade: far fewer false alarms, more misses. It's also Alibaba's benchmark on Alibaba's tool, though the dataset is on Hugging Face. If you want help reading numbers like these, see coding agent benchmarks explained.

The model didn't get smarter. It got fewer decisions to make, because plain code handles everything around it:

  1. Files: a filter drops binaries, your excludes, test files and vendor folders.
  2. Grouping: one quick model call sorts related files by name. Code caps each group at ten, and each group gets its own sub-agent, eight at a time.
  3. Rules: every file gets the rules for its type, from 64 rule files covering Go to Terraform.
  4. Review: the model reviews with tools to read whole files and search your code.
  5. Cleanup: a separate step matches each comment to its exact line, and another pass throws out clearly wrong comments.

I ran delegation mode, which needs no model key: your own agent does the review and Open Code Review hands it the plan. On a small demo repo, it picked five of six changed files, skipped the package lock and sorted them into four rule groups. I had planted a SQL injection, and the Java rules it handed back don't mention SQL injection. I didn't run the full review, so I can't say whether the model would catch it. There's also a GitHub Actions recipe that reviews every pull request and posts comments on the lines.

How I'd stack them

One agent builds, one agent reviews. The builder gets FastMCP for its own tools, Stitch Skills for design, last30days for research and ffmpeg-skill if you make video. Open Code Review sits on the other side of the pull request, checking with rules instead of vibes.

Key takeaways

  • If I install one tonight, it's Open Code Review in delegation mode, because it needs no new key. FastMCP comes next, the first time my agent writes the same script twice.
  • Open Code Review's gain came from structure, not a better model: code picks files, caps groups and matches rules, and the model only reviews.
  • ffmpeg-skill is built around verification: probe first, typed tools, measured output and a contact sheet the agent has to look at.
  • last30days works without keys, but expect a thinner brief than the README suggests.
  • Stitch Skills is the one I couldn't run. It needs the Stitch MCP server and your own API key, and it isn't moving fast.

FAQ

What is Alibaba's Open Code Review?

It's an open source (Apache 2.0) code reviewer from Alibaba. Plain code picks the files, caps groups at ten and matches 64 rule files by file type, the model does the review, and a separate step matches each comment to its exact line.

Is Open Code Review better than Claude Code at code review?

On Alibaba's own AACR-Bench (200 pull requests, same Claude Opus 4.6 model), 34% of Open Code Review's comments were real issues versus about 7% inside Claude Code, using about 385K tokens per review instead of about 5.7M. Claude Code still found more real issues overall, 435 versus 301.

What does FastMCP do?

FastMCP turns a normal Python function into an MCP tool with one decorator. The function name becomes the tool name, the type hints become the input schema, the docstring becomes the description, and every call is checked before your code runs.

Do these repos need API keys?

ffmpeg-skill needs none, only FFmpeg and Python on your machine. Open Code Review's delegation mode needs no extra key, and last30days works keyless for Reddit, Hacker News, Polymarket and GitHub. Stitch Skills needs the Stitch MCP server and your own Stitch API key.

Run your agents in Beckon

A workspace for your coding agents. Get more work done in less time.

Download BeckonSubscribe on YouTube
Full video transcript

0:00Alibaba open- sourced the code reviewer it says its developers used for 2 years and on Alibaba's own test the same claude model went from 7% of its comments being real issues to 34%. And the model didn't get any smarter. Alibaba just gave it fewer decisions to make. One repo gives your agent 42 FFmpeg tools. Google's gives it a designer. So in this video I'm counting down five trending repos that each hand your agent a new job. and the one I'd install tonight. I ran every one I could. No API keys. Let's jump right into it. So, where did these five come from? I pulled GitHub trending on Sunday night, the daily and the weekly charts for all languages and the big ones like

0:45Python, TypeScript, and Go. I only kept repos that hand your agent a new job. I skipped anything we've covered before and anything that's really a whole app for running agents. Four of these are on the weekly charts, so they're ranked by stars this week. The fifth only shows up on the daily chart, so it starts the board at number five, even though of the ones you can count, it gets by far the most downloads. Number five. Number five is Fast MCP from Prefect. 82 stars today on the Python chart and over 28,000 total. Job number one, your agent's toolmaker. So MCP is how your agent talks to tools. A server tells the agent, "Here's what I can do. Here's what each tool takes," and the agent calls it. Writing that by hand is a lot

1:32of boilerplate. With FastMPP, you write a normal Python function, put one decorator on top of it, and that's a tool. So, what does that one decorator actually do? It reads your function. The name becomes the tool name. The type hints become the input schema, and the dock string becomes the description your agent reads to decide when to use it. Then it checks every call before your code runs. So if the agent sends a word where you wanted a number, it gets bounced before it ever touches your function. So I tried it. I wrote a tiny server with one tool, count to-dos, that finds the to-do and fix me comments in a folder, 23 lines. Then fast MCP's inspect command printed the schema it built, the path, the file types with their defaults, which one is required. I

2:17didn't write a single line of that schema. And just for fun, I pointed it at the official MCP Python SDK. 31 to-do and fix me notes across 18 files. Prefect says fast MCP 1.0 got folded into that same SDK back in 2024 and that some version of it runs 70% of MCP servers. That's their number. The downloads you can check yourself. And Pippi stats shows 1.7 million in a single day, which is insane. So, if your agent keeps writing the same curl command over and over, give it a real tool instead. Want to make videos like this one with Claude Code? That's what I teach step by step inside Shipping School Premium. New lessons every single day, live calls every week, a full

3:02course library, and over 500 hours of replays. 263 builders are already in. Join free. Link below. Number four is Stitch Skills from Google Labs. 442 stars this week. Job number two, your agents designer. Stitch is Google's AI design tool. You describe a screen and it designs it. These skills connect your agent to it through the Stitch MCP server in Claude Code, Codeex, Cursor, Gemini CLI, and Anti-gravity. It comes as three plugins. Design makes new screens, edits them, makes variants, and can pull a design system out of code you already have. Build turns those screens into React components, React Native or a Vit dashboard and it helps with shad CN

3:47too. And utilities cleans up your prompt and runs a build loop. So how does design to code actually work? The React skill is strict. It downloads every screen's HTML and a screenshot and it has to actually look at the screenshot. Then it pulls the colors, the fonts, and the spacing out of the design into a style guide. Then it builds the components and if you say yes, it runs checks on them. And every step has a gate in the skill file in capital letters saying do not skip this, which honestly is how you have to talk to an agent sometimes. So the build loop is the fun one. It uses a baton file. Each run builds one page, then writes the prompt for the next page and hands it off. So this is where I'd slow down. All of this needs the Stitch MCP server and

4:33your own Stitch API key. and I didn't set that up for this video. So, no live run on this one, just the skill files. And the last commit was back in August, so it's trending, but it isn't moving fast. Google also says it's not an officially supported product. Number three is last 30 days from Matt Van Horn. 794 stars this week and almost 68,000 total. His GitHub says he co-founded the company that became Lyft. Job number three, your agents researcher. So you type slashlast30 days and a topic and it goes through Reddit X, YouTube, HackerNews, Poly Market, and the web for what people actually said in the last 30 days. Then it comes back with one brief. So what happens after you hit enter? First, your agent figures

5:19out who matters, the X handles, the subreddits, the GitHub repo. Then it searches every source at the same time. Then it scores everything by what real people engaged with. Upvotes, likes, comments, even polymarket odds. Then it merges the same story from different sites into one and writes the brief with the sources cited. And after that, your session knows all of it so you can keep asking questions. The readme says Reddit, HackerNews, Poly Market, and GitHub work with no keys at all. X and YouTube need some setup and it can use your browser login for X if you let it. I didn't. So I ran it on Claude Code Skills. no keys and the research part finished in 4.9 seconds. 15 hacker news stories, 12 polyarket markets, and three

6:04GitHub results all scored and grouped. Reddit got blocked from my network, so that was zero. And the brief says so instead of pretending, but most of what came back was general Claude Code news because the hacker news search dropped the word skills. So, it's fast and it's honest about what it missed. But the keyless version is thinner than the readme makes it sound. So, here's how I'd use it. Before I cover a repo, I want to know what people are actually running into with it, not what the readme says. That used to be 20 tabs. Now it's one command. Number two is ffmpeg skill. 1110 stars this week. Job number four, your agent's video editor. It's 42 FFmpeg tools for cloud code, cursor, and codecs, plus an MCP server. No cloud, no API keys, just FFmpeg and

6:52Python on your own machine. It's from Shogo Kajayyama, a video and streaming engineer. And fun fact, Claude has 53 commits in this repo. So why not just let your agent write ffmpeg commands? Because it guesses. It guesses the frame rate, picks a codec the file can't hold, and tells you done without ever opening the result. And a bad ffmpeg command will often still finish with no error. So how does this fix it? Every job starts with probe, which measures the real file. Then the edit runs as a typed tool, never a raw shell command. Then it measures its own output, the length, the loudness, and if the picture changed, it makes a contact sheet, and the agent has to look at it. The report isn't done until it names that image. And the schema your agent sees is generated from

7:39each script's own arguments, so the tool and its description can't drift apart. If they do, the build fails. Over MCP, it only lists the core 12 tools, so it doesn't eat your context. But all 42 are there by name. So I ran it the way an agent would on a 45se secondond test clip. Probe measured it. Cut kept two 10-second pieces with no re-encode. Fit made it 15 seconds and vertical. Caption added the text and loudness took it from minus20.7 to minus14. Then check ran 12 checks for shorts. Zero failed. And look made the contact sheet. And you can see the final file. 15 seconds 1080 by 1920 30 frames a second. They also test it with agents. In one round, Sonnet ran a 100 prompts in 17 languages and answered

8:26in the right language every time. And the greater caught zero fake successes. Their own test and the misses are written up in the repo, too. And number one with 2,238 stars this week and over 52,000 total is open code review from Alibaba. Job number five, the code reviewer. Alibaba says tens of thousands of their developers used it. It's open source Apache 2.0. So, like I said, it's the same model both times. Claude Opus 4.6 on 200 real poll requests. Inside Claude Code, it left 5,980 comments and only about 7% were real issues. Inside open code review, 889 comments and 34% were real. It used

9:12about 385,000 tokens a review instead of about 5.7 million and it took about a minute and a half instead of 13. Now, to be fair, Claude Code actually found more real issues. 435 versus 301. That's the trade they picked. Way fewer false alarms, but more misses. And this is Alibaba's benchmark on Alibaba's tool. So, take that for what you will, but they put the data set up on HuggingFace. So, how does Alibaba pull that off? The model makes fewer decisions. Plain code handles everything around the model that can't go wrong. First, which files? A filter drops binaries, your excludes, test files, and vendor folders. Second, grouping. One quick model call looks at just the file names and sorts related

9:58files together. Plain code caps each group at 10, and each group gets its own sub aent with its own context, eight at a time. Third, rules. Every file gets the rules for its type and there are 64 rule files in there from Go to Terraform. Then the model does the actual review with tools to read whole files and search your code. And then a separate step matches each comment back to the exact line and another pass throws out comments that are clearly wrong before any of it reaches you. Now, I didn't put a key in, so I ran the part that needs no model. It's called delegation mode. Your own agent does the review with its own subscription and Open Code Review just hands it the plan. On a little demo repo I made, it picked five of the six changed files, skipped

10:43the package lock and sorted them into four rule groups. Java, the two language files together, package.json, and Typescript. One thing I noticed, I put a SQL injection in there on purpose, and the Java rules it handed back don't mention SQL injection at all. I didn't run the full review, so I can't say if the model would catch it. And there's a GitHub actions recipe that reviews every pull request and posts the comments right on the lines. So, how do these fit together? One agent builds, one agent reviews. The builder gets fast MCP for its own tools. Stitch for design, last 30 days for research, and FFmpeg Skill. If you make any kind of video, then open code review sits on the other side of the pull request, and it's checking with rules, not vibes. So, if I'm installing

11:30one tonight, it's open code review in delegation mode because it needs no new key. Then fast MCP, the next time my agent writes the same script twice. So, which of these jobs would you actually trust your agent with first? Tell me in the comments. And if you end up installing one of these, hit like so I know to keep this series going. Have a blessed day.