more codemode fun, doing a large mysql operation and the agent wanted to do it 50K rows at a time
it writes a single codemode script which runs it in a loop
and then you visualize it step by step even though it's code
it writes a single codemode script which runs it in a loop
and then you visualize it step by step even though it's code
September 18, 2026 at 1:03 AM
more codemode fun, doing a large mysql operation and the agent wanted to do it 50K rows at a time
it writes a single codemode script which runs it in a loop
and then you visualize it step by step even though it's code
it writes a single codemode script which runs it in a loop
and then you visualize it step by step even though it's code
if mcp servers implement their own codemode
and then your agent native supports codemode
now everything is going through an extra useless layer
if you insist on doing this plz provide a more raw mcp endpoint
and then your agent native supports codemode
now everything is going through an extra useless layer
if you insist on doing this plz provide a more raw mcp endpoint
July 20, 2026 at 1:27 AM
if mcp servers implement their own codemode
and then your agent native supports codemode
now everything is going through an extra useless layer
if you insist on doing this plz provide a more raw mcp endpoint
and then your agent native supports codemode
now everything is going through an extra useless layer
if you insist on doing this plz provide a more raw mcp endpoint
opencode2 has tool search and codemode built in
this means it can handle infinite mcps, custom tools, apis, etc
and it's the simplest implementation we've seen
this means it can handle infinite mcps, custom tools, apis, etc
and it's the simplest implementation we've seen
August 7, 2026 at 11:10 PM
opencode2 has tool search and codemode built in
this means it can handle infinite mcps, custom tools, apis, etc
and it's the simplest implementation we've seen
this means it can handle infinite mcps, custom tools, apis, etc
and it's the simplest implementation we've seen
fiend - build 3D stuff with your agent in the browser
connect the mcp server, ask your agent to make a new scene and get a link. you can point out areas for feedback and prompt it further
works beautifully with opencode2's codemode
connect the mcp server, ask your agent to make a new scene and get a link. you can point out areas for feedback and prompt it further
works beautifully with opencode2's codemode
September 9, 2026 at 3:30 AM
fiend - build 3D stuff with your agent in the browser
connect the mcp server, ask your agent to make a new scene and get a link. you can point out areas for feedback and prompt it further
works beautifully with opencode2's codemode
connect the mcp server, ask your agent to make a new scene and get a link. you can point out areas for feedback and prompt it further
works beautifully with opencode2's codemode
Just as MCP seems to be everywhere, I'm sneaking out the back door. I really only use it for things that need auth or UI. Skills and codemode do most everything and agents handle them more naturally
March 13, 2026 at 12:42 AM
Just as MCP seems to be everywhere, I'm sneaking out the back door. I really only use it for things that need auth or UI. Skills and codemode do most everything and agents handle them more naturally
our anti fraud system gets access to honeycomb, stripe, and database to correlates things and find very obvious fraud
the models are incredible at doing this via codemode. the whole script didn't even fit in the screenshot
composing these tools together to get what it needs
the models are incredible at doing this via codemode. the whole script didn't even fit in the screenshot
composing these tools together to get what it needs
August 26, 2026 at 1:30 PM
our anti fraud system gets access to honeycomb, stripe, and database to correlates things and find very obvious fraud
the models are incredible at doing this via codemode. the whole script didn't even fit in the screenshot
composing these tools together to get what it needs
the models are incredible at doing this via codemode. the whole script didn't even fit in the screenshot
composing these tools together to get what it needs
codemode is so good man it just did a giant analysis to answer a question i had in one go
this is pretty much a non-negotiable thing at this point
this is pretty much a non-negotiable thing at this point
August 25, 2026 at 9:51 PM
codemode is so good man it just did a giant analysis to answer a question i had in one go
this is pretty much a non-negotiable thing at this point
this is pretty much a non-negotiable thing at this point
now that we have codemode you can add as many MCPs as you want without cost
i still don't add a ton up front but occasionally i'll come across another good use case
and over time my agent gets more and more capable - esp with project specific ones
i still don't add a ton up front but occasionally i'll come across another good use case
and over time my agent gets more and more capable - esp with project specific ones
September 11, 2026 at 6:42 PM
now that we have codemode you can add as many MCPs as you want without cost
i still don't add a ton up front but occasionally i'll come across another good use case
and over time my agent gets more and more capable - esp with project specific ones
i still don't add a ton up front but occasionally i'll come across another good use case
and over time my agent gets more and more capable - esp with project specific ones
still, in opencode2 we implemented codemode in a way that allows us to render decent ui and still prompt for permissions even when the model is just lobbing over a hunk of code
but if astra really wants to write raw python, should we even be fighting that and making it use our own thing? 2/3
but if astra really wants to write raw python, should we even be fighting that and making it use our own thing? 2/3
September 6, 2026 at 3:36 PM
still, in opencode2 we implemented codemode in a way that allows us to render decent ui and still prompt for permissions even when the model is just lobbing over a hunk of code
but if astra really wants to write raw python, should we even be fighting that and making it use our own thing? 2/3
but if astra really wants to write raw python, should we even be fighting that and making it use our own thing? 2/3
my kingdom for a wasm codemode harness with a granular permission model + memgpt
June 3, 2026 at 12:03 AM
my kingdom for a wasm codemode harness with a granular permission model + memgpt
Still think we should deprecate MCP. It was built around the idea of tool calls, which turned out to be a mistake IMHO. And codemode is a limited hack.
Models mostly output Bash scripts anyway and they should probably *only* output Bash scripts.
Structured output can be produced from bash scripts.
Models mostly output Bash scripts anyway and they should probably *only* output Bash scripts.
Structured output can be produced from bash scripts.
i complained a lot about the mcp spec in the past
but it's getting very good now
but it's getting very good now
August 26, 2026 at 1:56 AM
Still think we should deprecate MCP. It was built around the idea of tool calls, which turned out to be a mistake IMHO. And codemode is a limited hack.
Models mostly output Bash scripts anyway and they should probably *only* output Bash scripts.
Structured output can be produced from bash scripts.
Models mostly output Bash scripts anyway and they should probably *only* output Bash scripts.
Structured output can be produced from bash scripts.
A few weeks ago our founder announced Monty, a minimal, secure Python interpreter written in Rust, for running code written by AI agents. People got excited. @sa.muelcolvin.com just wrote up the full story on why he's excited about it and you should be too.
pydantic.dev/articles/pyd...
pydantic.dev/articles/pyd...
Pydantic Monty: A Minimal Python Sandbox for AI Agents | Pydantic
Meet Monty — a secure, minimal Python interpreter written in Rust for running AI-generated code. Learn where it sits between tool calling and full computer use, and why CodeMode matters for AI agents.
pydantic.dev
February 27, 2026 at 5:01 PM
A few weeks ago our founder announced Monty, a minimal, secure Python interpreter written in Rust, for running code written by AI agents. People got excited. @sa.muelcolvin.com just wrote up the full story on why he's excited about it and you should be too.
pydantic.dev/articles/pyd...
pydantic.dev/articles/pyd...
Code Mode for MCP is truly revolutionary for the right use cases! Check out this blog post for a prime example of an optimal Code Mode MCP use case!
#MCP #AIDrivenDevelopment #DevTools #ContextEngineering #CodeMode #ClaudeCode #Codex #Gemini #OpenCode #KiloCode #Cursor #CodeGraph #devbubble
#MCP #AIDrivenDevelopment #DevTools #ContextEngineering #CodeMode #ClaudeCode #Codex #Gemini #OpenCode #KiloCode #Cursor #CodeGraph #devbubble
We replaced 11 MCP tool definitions with one that lets the LLM write code. Parallel queries, custom response shaping, zero intermediate payloads in context. Full credit to @cloudflare.social for pioneering the pattern! Here's what we learned adopting it: blog.constellationdev.io/2026/04/15/code-mode
Code Mode: Rethinking MCP | Constellation Blog
Traditional MCP tool calling doesn't scale. Learn how Code Mode lets AI agents compose queries as code, cutting token overhead and unlocking dynamic multi-step analysis in a single invocation.
blog.constellationdev.io
April 15, 2026 at 5:09 PM
Code Mode for MCP is truly revolutionary for the right use cases! Check out this blog post for a prime example of an optimal Code Mode MCP use case!
#MCP #AIDrivenDevelopment #DevTools #ContextEngineering #CodeMode #ClaudeCode #Codex #Gemini #OpenCode #KiloCode #Cursor #CodeGraph #devbubble
#MCP #AIDrivenDevelopment #DevTools #ContextEngineering #CodeMode #ClaudeCode #Codex #Gemini #OpenCode #KiloCode #Cursor #CodeGraph #devbubble
I recently was building with TanStack AI just had to interview @jherr.dev about its upcoming codemode features and how chat apps are the "hello world" of agent development! Check it out: www.youtube.com/watch?v=DKkE...
TanStack AI and the future of agentic tooling
Jack Herrington has been writing JavaScript for 40 years. As the Blue Collar Coder on YouTube and a core contributor to TanStack at Netlify, he's now building TanStack AI — a type-safe,…
www.youtube.com
March 17, 2026 at 1:35 PM
I recently was building with TanStack AI just had to interview @jherr.dev about its upcoming codemode features and how chat apps are the "hello world" of agent development! Check it out: www.youtube.com/watch?v=DKkE...
just want to thank @kentonvarda.com @threepointone.bsky.social and the cloudflare ai teams for codemode 🙏
i've been meaning to try it for some time, and paired with sonnet it's just opened up really cool stuff for us we couldn't do with major providers...
i've been meaning to try it for some time, and paired with sonnet it's just opened up really cool stuff for us we couldn't do with major providers...
March 9, 2026 at 8:55 AM
just want to thank @kentonvarda.com @threepointone.bsky.social and the cloudflare ai teams for codemode 🙏
i've been meaning to try it for some time, and paired with sonnet it's just opened up really cool stuff for us we couldn't do with major providers...
i've been meaning to try it for some time, and paired with sonnet it's just opened up really cool stuff for us we couldn't do with major providers...
Datalayer is writing a paper on cost-effective LLM agents (70–92% token reduction via codemode).
Seeking co-authors with real-world agent data.Read & apply → datalayer.ai/research/con...
#LLMAgents #AIEngineering #OpenScience
Seeking co-authors with real-world agent data.Read & apply → datalayer.ai/research/con...
#LLMAgents #AIEngineering #OpenScience
datalayer.ai
March 3, 2026 at 2:53 PM
Datalayer is writing a paper on cost-effective LLM agents (70–92% token reduction via codemode).
Seeking co-authors with real-world agent data.Read & apply → datalayer.ai/research/con...
#LLMAgents #AIEngineering #OpenScience
Seeking co-authors with real-world agent data.Read & apply → datalayer.ai/research/con...
#LLMAgents #AIEngineering #OpenScience
i think it’s still underrated for a few reasons:
- many people don’t what the point of DSLs are
- codemode/full computer use is the current meta and north star
- most people read the bitter lesson to mean “let the model figure out everything from scratch”
- designing DSLs is hard
- many people don’t what the point of DSLs are
- codemode/full computer use is the current meta and north star
- most people read the bitter lesson to mean “let the model figure out everything from scratch”
- designing DSLs is hard
June 15, 2026 at 9:18 AM
i think it’s still underrated for a few reasons:
- many people don’t what the point of DSLs are
- codemode/full computer use is the current meta and north star
- most people read the bitter lesson to mean “let the model figure out everything from scratch”
- designing DSLs is hard
- many people don’t what the point of DSLs are
- codemode/full computer use is the current meta and north star
- most people read the bitter lesson to mean “let the model figure out everything from scratch”
- designing DSLs is hard
Proxy #MCP servers via @cloudflare.social #AI Controls. Instead of loading every tool into context, #CodeMode turns an agent's tool into an #API. #Agent chain & filter across dozens of tools in one pass, without a round-trip to model for every call. Cut #token #costs by ~93%
#LLM #Agents
#LLM #Agents
Cloudflare Dashboard | Manage Your Account
Log in to the Cloudflare dashboard. Make your websites, apps, and networks fast and secure. Build modern apps on our developer platform.
dash.cloudflare.com
July 24, 2026 at 12:25 AM
https://github.com/jx-codes/codemode-mcp
LLMがツール呼び出しよりもコード作成に適しているという考えに基づいたツールです。
LLMはTypeScript/JavaScriptコードを記述し、HTTPプロキシを介してMCPサーバーにアクセスします。
これにより、複雑なツール連携の利点を活用しつつ、LLMが得意なコード作成に集中できます。
LLMがツール呼び出しよりもコード作成に適しているという考えに基づいたツールです。
LLMはTypeScript/JavaScriptコードを記述し、HTTPプロキシを介してMCPサーバーにアクセスします。
これにより、複雑なツール連携の利点を活用しつつ、LLMが得意なコード作成に集中できます。
GitHub - jx-codes/codemode-mcp: An implementation of the Codemode workflow detailed by Cloudflare.
An implementation of the Codemode workflow detailed by Cloudflare. - jx-codes/codemode-mcp
github.com
September 29, 2025 at 11:52 AM
https://github.com/jx-codes/codemode-mcp
LLMがツール呼び出しよりもコード作成に適しているという考えに基づいたツールです。
LLMはTypeScript/JavaScriptコードを記述し、HTTPプロキシを介してMCPサーバーにアクセスします。
これにより、複雑なツール連携の利点を活用しつつ、LLMが得意なコード作成に集中できます。
LLMがツール呼び出しよりもコード作成に適しているという考えに基づいたツールです。
LLMはTypeScript/JavaScriptコードを記述し、HTTPプロキシを介してMCPサーバーにアクセスします。
これにより、複雑なツール連携の利点を活用しつつ、LLMが得意なコード作成に集中できます。
get my DIY/nights-and-weekends AI lab functional enough to start producing meaningful R&D
(initial side-quests: golden-gate-gemma-J-space, virtual tools + git-ified codemode, RegimeBenchmaxxing via RLAIF, Lean-4-fun-and-profit, SOTA full-press AI Diplomacy, CUDA-specialist SLMs)
(initial side-quests: golden-gate-gemma-J-space, virtual tools + git-ified codemode, RegimeBenchmaxxing via RLAIF, Lean-4-fun-and-profit, SOTA full-press AI Diplomacy, CUDA-specialist SLMs)
September 6, 2026 at 8:03 PM
get my DIY/nights-and-weekends AI lab functional enough to start producing meaningful R&D
(initial side-quests: golden-gate-gemma-J-space, virtual tools + git-ified codemode, RegimeBenchmaxxing via RLAIF, Lean-4-fun-and-profit, SOTA full-press AI Diplomacy, CUDA-specialist SLMs)
(initial side-quests: golden-gate-gemma-J-space, virtual tools + git-ified codemode, RegimeBenchmaxxing via RLAIF, Lean-4-fun-and-profit, SOTA full-press AI Diplomacy, CUDA-specialist SLMs)
The Webwright write up is great. Talks up very signficiant wins for #CodeMode agenting; the ability to issue multiple steps at once seems like a very big win! And everything is written down in the end!
www.microsoft.com/en-us/resear...
www.microsoft.com/en-us/resear...
May 26, 2026 at 4:12 PM
The Webwright write up is great. Talks up very signficiant wins for #CodeMode agenting; the ability to issue multiple steps at once seems like a very big win! And everything is written down in the end!
www.microsoft.com/en-us/resear...
www.microsoft.com/en-us/resear...
Dynamic Worker Loader now in public beta 👀 blog.cloudflare.com/dynamic-work...
I'm a little sad this is focusing so much on codemode / MCP / Agents. IMO the feature is *much* more interesting than that. If you need to run untrusted code securely and cheaply, you need something like DWL.
I'm a little sad this is focusing so much on codemode / MCP / Agents. IMO the feature is *much* more interesting than that. If you need to run untrusted code securely and cheaply, you need something like DWL.
Sandboxing AI agents, 100x faster
We’re introducing Dynamic Workers, which allow you to execute AI-generated code in secure, lightweight isolates. This approach is 100 times faster than traditional containers, enabling millisecond sta...
blog.cloudflare.com
March 24, 2026 at 7:05 PM
Dynamic Worker Loader now in public beta 👀 blog.cloudflare.com/dynamic-work...
I'm a little sad this is focusing so much on codemode / MCP / Agents. IMO the feature is *much* more interesting than that. If you need to run untrusted code securely and cheaply, you need something like DWL.
I'm a little sad this is focusing so much on codemode / MCP / Agents. IMO the feature is *much* more interesting than that. If you need to run untrusted code securely and cheaply, you need something like DWL.
🧠 Context engineering > prompt engineering.
We compare CodeMode vs no CodeMode with reproducible GitHub Actions workflows. Less context → fewer tokens → lower cost.
🎥 Demo: www.loom.com/share/1edbbe...
Report deep dive next.
We compare CodeMode vs no CodeMode with reproducible GitHub Actions workflows. Less context → fewer tokens → lower cost.
🎥 Demo: www.loom.com/share/1edbbe...
Report deep dive next.
Datalayer ☰ AI Agents for Data Analysis - 17 July 2026
Use Loom to record quick videos of your screen and cam. Explain anything clearly and easily – and skip the meeting. An essential tool for hybrid workplaces.
www.loom.com
July 17, 2026 at 5:14 PM
🧠 Context engineering > prompt engineering.
We compare CodeMode vs no CodeMode with reproducible GitHub Actions workflows. Less context → fewer tokens → lower cost.
🎥 Demo: www.loom.com/share/1edbbe...
Report deep dive next.
We compare CodeMode vs no CodeMode with reproducible GitHub Actions workflows. Less context → fewer tokens → lower cost.
🎥 Demo: www.loom.com/share/1edbbe...
Report deep dive next.
**The cheapest token is the one you never send.**
Most AI agents keep sending the same context back to the LLM.
CodeMode keeps execution inside a **Code Sandbox**, sending only what's needed for the next decision.
Fewer tokens. Lower cost. Faster agents.
datalayer.ai/guide
#AI #AIAgents #LLM
Most AI agents keep sending the same context back to the LLM.
CodeMode keeps execution inside a **Code Sandbox**, sending only what's needed for the next decision.
Fewer tokens. Lower cost. Faster agents.
datalayer.ai/guide
#AI #AIAgents #LLM
Datalayer ☰ AI Agents for Data Analysis
Datalayer ☰ AI Agents for Data Analysis
datalayer.ai
July 31, 2026 at 3:29 PM
**The cheapest token is the one you never send.**
Most AI agents keep sending the same context back to the LLM.
CodeMode keeps execution inside a **Code Sandbox**, sending only what's needed for the next decision.
Fewer tokens. Lower cost. Faster agents.
datalayer.ai/guide
#AI #AIAgents #LLM
Most AI agents keep sending the same context back to the LLM.
CodeMode keeps execution inside a **Code Sandbox**, sending only what's needed for the next decision.
Fewer tokens. Lower cost. Faster agents.
datalayer.ai/guide
#AI #AIAgents #LLM