I built a RAG and knowledge graph agent that runs locally
Claw-Coder is an AI agent that runs locally on your laptop and has access to powerful tools instead of configuring claude or codex to use a local model just use claw-coder. Why was claw-coder created? Answer: To solve the problem of privacy and security. When you use an agent that is configured with a cloud model like codex, cursor, Claude etc. You are not just getting the agent but you are giving up your codebase to train an llm which is a bit concerning and this reduces trust in the technology called AI but now another problem comes in performance when you switch to a local model that is…
What it does
In the maker’s words, at launch
Claw-Coder is an AI agent that runs locally on your laptop and has access to powerful tools instead of configuring claude or codex to use a local model just use claw-coder. Why was claw-coder created? Answer: To solve the problem of privacy and security. When you use an agent that is configured with a cloud model like codex, cursor, Claude etc. You are not just getting the agent but you are giving up your codebase to train an llm which is a bit concerning and this reduces trust in the technology called AI but now another problem comes in performance when you switch to a local model that is not made for that workflow you lose performance, speed, and it becomes really a tradeoff so that's where claw-coder comes in it not only runs on your machine but all the code, rag, knowledge graph etc info is kept local making the privacy problem solved but now what about performance. Performance: Local llms are not built to do the cool things cloud models do because the model sizes are not even capable of building real apps like the 8b models, 13b, even 1b so the solution I came up with was to give these small models access to tools and features that make it actually work well in coding performance. So what does claw-coder have access to: A knowledge graph: A knowledge graph is an interconnected network of real-world entities—such as people, places, concepts, or events—and the relationships between them. It organizes information into a readable web of meaning rather than static lists, allowing both humans and AI to understand context. So how does this help an AI, it gives the AI the ability to tell relationships between code in your codebase, a cloned unknown repo and so forth this increases performance of local llms by far in coding tasks and reasoning abilities. Rag: We have ever had of RAG at some point but there is a catch the context window of local llms can't bear large codebases and repos so RAG isn't optional by storing vectors in a vector store you enable the AI to actually know what the code means and what each piece does to the other letting you load millions of lines into the vector store without blowing up the context window. Tools: So we have discussed the tiny but powerful ways to improve local llm performance but an agent to be an agent it needs to take action this is where exposing tools to the local llm helps so what tools have been implemented into claw-coder. 1. search_tool This enables the ai agent to actually search up to date info so that it doesnt hallucinate on info it doesn't know which is common in local llms. 2. Docker execution This agent has a special folder called workspace where it does its work without destroying your desktop but this is not enough to protect your desktop from being destroyed by cheap code so this is where docker comes in I have implemented docker containers of various languages where the agent can validate its own code this is powerful because all llms not only local ones generate code they can't even confirm works because they are just powerful predictors so enabling it to run its code can surprisingly increase the usefulness of the generated code because it now knows it works or not even for html and css the ai agent has been given a helpful vision llm to actually explain what rendered in the browser. This is the surprising power of giving an llm a docker execution tool. We have looked at a lot of how claw-coder is different enabling local llms to actually do real work. But how do you actually try it out yourself: Claw-coder is closed source because it is going through heavy testing but that doesnt kill transparency and since we are testing it doesn't stop people from trying it on real codebase and giving feedback to get started use: brew tap gabriel-c70/claw then brew install claw-coder
Does the same job
all alternatives →- IBI built a powerful RAG and knowledge graph agent that runs locallyMay 2026 · ▲5
Claw-Coder is an AI agent that runs locally on your laptop and has access to powerful tools instead of configuring claude or codex to use a local model just use claw-coder. Why was claw-coder created? Answer: To solve the problem of privacy and security. When you use an agent that is configured with a cloud model like codex, cursor, Claude etc. You are not just getting the agent but you are giving up your codebase to train an llm which is a bit concerning and this reduces trust in the technology called AI but now another problem comes in performance when you switch to a local model that is…
More ai this month
the category →
I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15). The idea is basically GitHub Copilot or Tabnine, except instead of prompting it with code, you prompt it by playing a few notes on a MIDI piano. The model then continues what you played, entirely on-device. The app is free if anyone wants to try it. Happy to answer questions about the model, training, Core ML, or the many things that didn't work.
AI · 17d ago · simedw.com
Astute▲585Automate your B2B brand going viral, with new media creators
AI · 18d ago · company-app.joinastute.com
Hey HN, Henry from Cactus here! We previously released Cactus Needle, a 14MB agentic LLM for tool call, device use, and structured extraction for phones, wearables, smart homes, small robots and microcontrollers. We got really great feedback here, and have now incorporated the suggestions to release Needle 2. The whole model is a single 14MB binary that runs a full session in 28MB of RAM; 45m parameters at 2bit compression. Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, sits between 400-1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges…
AI · 26d ago · cactuscompute.com
Launched alongside, May 2026
the whole month →
Parallel agents, diff reviewer, and multi-model comparisons
Dev tools · May 2026 · kilo.ai
- NW
Hey HN, Henry here from Cactus. We open-sourced Needle, a 26M parameter function-calling (tool use) model. It runs at 6000 tok/s prefill and 1200 tok/s decode on consumer devices. We were always frustrated by the little effort made towards building agentic models that run on budget phones, so we conducted investigations that led to an observation: agentic experiences are built upon tool calling, and massive models are overkill for it. Tool calling is fundamentally retrieval-and-assembly (match query to tool name, extract argument values, emit JSON), not reasoning. Cross-attention…
Life & fun · May 2026 · github.com
- FM
Dev tools · May 2026 · github.com