Correct, this is like giving the LLM full specs on your GraphQL endpoint, and telling it have fun, get only what you need.
Except it can use Typescript to manipulate the data mid stream.
So in reality this is MCP Middleware + MCP Orchestration + Tool Call Efficiency.
It will save tokens as well due to only returning what it needs, but more so just think of going from 10 consecutive tool calls, to 1 call that gives you everything you need and nothing you don't, with 1/10th the time taken to accomplish.
In which case your likely wrong, people do want it, and AI will be very good at orchestrating simple patterns.
CF definitely has a vested interest.. The problem for them now that I see is that THEY DIDN"T ACTUALLY LAUNCH IT... but did describe what it is/does in complete detail.
Now there are gonna be dozens of non CF locked clones, just like the one OP linked.
I came here looking for partial solutions to fill gaps in our own attempts at CodeMode service. However I'll give you some hope ;)
> Later I want to see what'd happen if you give the LLM a repo of sorts to store useful snippets and functions with comments for later use. So the LLM itself would save workflows, be able to import them into the Deno environment and chain those together.
OMG this is the first thing you should do. We have similar now and its freaking amazing. Just discussing yesterday how I can't remember it going off the rails since implementing automem last week even.
Best thing it does, fully recaps all your daily accomplishments, across all platforms (Claude Code, Claude Desktop, ChatGPT, Cursor).
Called Automem by a friend of my (Jack Arturo), currently closed-source, though I'm sure you could reverse engineer it enough.
- its a hosted stack of FalkorDB + QDrant
- has endpoints for creating/retrieving memories
- embeds stuff using ChatGPT models
- Uses Graph nodes for relating memories together
- Has a dream/sleeping phase which degrades long term memory relevant, finds and tracks patterns and more.
- Has an MCP which connects any AI directly to memory
- Automated hooks which record memory queues on commit, deploy, learning moments
- Automated storing of all queued memories on chat end.
- A lot more magic under the hood too.
So in reality you get a near biological memory, useful by any MCP agent. To be fair Jack has about a 2 month head start on the rest of us with this idea haha.
--
The setup were building will be an always running setup, so it also has a scheduling runtime in Node that uses MD files to create automatable workflows, some uses agents, some just run bash. They can call mcps, tools, run commands, log output, use automem etc, all in human readable text.
PS Keep up the great work on your codemode service, got some great ideas from yours to incorporate to ours that should resolve the one or 2 issues we had outstanding. Will share if I get it working, https://github.com/danieliser/code-mode if it gets any where
Oh this sounds so awesome. The memory system sounds very cool. I could imagine decaying memories in the graph, their initial salience being N then as it gets referenced more and more N goes up (more edges?), keep a decay function over time?
Basically instead of mcp servers you write typescript files that are parsed to generate a typed client, these are executed in one deno sandbox, and the LLM code gets that typed client and its scripts are run in a sandbox with only net allowed.
Been having some fun testing it out today.
If you have time to take a look I would be curious to hear what you think.
Except it can use Typescript to manipulate the data mid stream.
So in reality this is MCP Middleware + MCP Orchestration + Tool Call Efficiency.
It will save tokens as well due to only returning what it needs, but more so just think of going from 10 consecutive tool calls, to 1 call that gives you everything you need and nothing you don't, with 1/10th the time taken to accomplish.