I just learned you can run Claude Code locally with Ollama.
Ollama 0.19 (preview, released yesterday) is now powered by Apple's MLX framework β and one thing that caught my attention as a Claude Code user: Ollama now reuses its cache across conversations, meaning less memory overhead and more cache hits when using a shared system prompt with tools like Claude Code.
That's a meaningful improvement for agentic workflows.
The setup is a single command:
Whether you're working in an air-gapped environment or just tired of API costs β local coding agents are getting genuinely viable.
(Requires a Mac with 32GB+ unified memory)
https://ollama.com/blog/mlx
Ollama 0.19 (preview, released yesterday) is now powered by Apple's MLX framework β and one thing that caught my attention as a Claude Code user: Ollama now reuses its cache across conversations, meaning less memory overhead and more cache hits when using a shared system prompt with tools like Claude Code.
That's a meaningful improvement for agentic workflows.
The setup is a single command:
ollama launch claude --model qwen3.5:35b-a3b-coding-nvfp4
Whether you're working in an air-gapped environment or just tired of API costs β local coding agents are getting genuinely viable.
(Requires a Mac with 32GB+ unified memory)
https://ollama.com/blog/mlx
Ollama
Ollama is now powered by MLX on Apple Silicon in preview Β· Ollama Blog
Today, we're previewing the fastest way to run Ollama on Apple silicon, powered by MLX, Apple's machine learning framework.
π5
Karpathy posted his raw/ folder workflow for LLM context β a simple idea: dump your notes, PDFs, papers into one folder and feed it to the LLM.
Someone built a full system on top of it for Claude Code.
The architecture is interesting:
β’ No RAG, no vector DB β just a smart index.md the LLM navigates
β’ Works at personal scale (50β500 articles) better than cosine similarity
β’ Runs on your existing Claude subscription, no extra API costs
https://github.com/coleam00/claude-memory-compiler
Someone built a full system on top of it for Claude Code.
claude-memory-compiler hooks into your Claude Code sessions and automatically compiles them into a structured knowledge base. Every decision, pattern, lesson β extracted and cross-referenced into markdown articles.The architecture is interesting:
β’ No RAG, no vector DB β just a smart index.md the LLM navigates
β’ Works at personal scale (50β500 articles) better than cosine similarity
β’ Runs on your existing Claude subscription, no extra API costs
https://github.com/coleam00/claude-memory-compiler
GitHub
GitHub - coleam00/claude-memory-compiler: Give Claude Code a memory that evolves with your codebase. Hooks automatically captureβ¦
Give Claude Code a memory that evolves with your codebase. Hooks automatically capture sessions, the Claude Agent SDK extracts key decisions and lessons, and an LLM compiler organizes everything in...
β€2π1
A drop-in Apache Spark replacement written in Rust, unifying batch processing, stream processing, and compute-intensive Al workloads.
https://github.com/lakehq/sail
https://github.com/lakehq/sail
GitHub
GitHub - lakehq/sail: Drop-in Apache Spark replacement written in Rust, unifying batch processing, stream processing, and computeβ¦
Drop-in Apache Spark replacement written in Rust, unifying batch processing, stream processing, and compute-intensive AI workloads. - lakehq/sail
π2
SLayer - an open-source semantic layer for AI agents.
The usual argument for giving agents more context is that it helps them write better SQL. SLayer takes a different angle: the problem isn't that agents write bad SQL, they're generally fine at that, it's that they don't know which SQL answers the question, so they guess, and the guess looks just as confident as a correct one. Same question, different query, sometimes a different number.
Instead of feeding the agent more context and hoping it reasons its way to the right query, SLayer gives it a governed set of pre-defined metrics to call. The agent still picks what it needs, but it never decides what "revenue" means or which tables to join, so the same question always returns the same, traceable answer.
Built by the team behind Motley, running as a library, CLI, or MCP/REST interface, no separate service required.
GitHub: github.com/MotleyAI/slayer
X: x.com/usemotley
Website: motley.ai/slayer
The usual argument for giving agents more context is that it helps them write better SQL. SLayer takes a different angle: the problem isn't that agents write bad SQL, they're generally fine at that, it's that they don't know which SQL answers the question, so they guess, and the guess looks just as confident as a correct one. Same question, different query, sometimes a different number.
Instead of feeding the agent more context and hoping it reasons its way to the right query, SLayer gives it a governed set of pre-defined metrics to call. The agent still picks what it needs, but it never decides what "revenue" means or which tables to join, so the same question always returns the same, traceable answer.
Built by the team behind Motley, running as a library, CLI, or MCP/REST interface, no separate service required.
GitHub: github.com/MotleyAI/slayer
X: x.com/usemotley
Website: motley.ai/slayer
π8π₯8β€3π2
Dagu: Local-first workflow orchestration that just works
https://dagu.sh/
https://dagu.sh/
Dagu
Dagu: Local-first workflow orchestration that just works
Local-first workflow orchestration in declarative YAML. One open-source binary with schedules, retries, human tasks, logs, and a Web UI. No database, no framework. Airflow alternative.
π₯2π1
AI Engineering from Scratch | Free AI Engineering Course
https://aiengineeringfromscratch.com/
https://aiengineeringfromscratch.com/
Aiengineeringfromscratch
AI Engineering from Scratch | Free AI Engineering Course
542 lessons. 20 phases. Write the backprop, the tokenizer, the attention mechanism, and the agent loop by hand.
β€3π2
Nice to see a lot of parallels with my PyData talk
- https://www.getdbt.com/blog/aws-dbt-semantic-layer-ai-agent-revenue-accuracy
- https://www.getdbt.com/blog/dbt-semantic-layer-ai-agent-adoption
- https://www.getdbt.com/blog/aws-dbt-semantic-layer-ai-agent-revenue-accuracy
- https://www.getdbt.com/blog/dbt-semantic-layer-ai-agent-adoption
dbt Labs
Why an AI agent's revenue query was $3.5 million wrong, and how AWS fixed it with dbt Semantic Layer | dbt Labs
AWS shows how pairing dbt Semantic Layer with the Amazon Redshift MCP Server helps AI agents use governed metric definitions.
β€3
Edelbyte left stealth with $70M (Sequoia, Quiet Capital) at a ~$500M valuation. The four founders are ex-Azure Kusto (Azure Data Explorer) people, building agent-oriented search and analytics for hundreds of petabytes. It will compete with Snowflake, Databricks and ClickHouse. https://siliconangle.com/?p=853911
SiliconANGLE
Edelbyte reels in $70M for its multipurpose data analytics platform
Edelbyte reels in $70M for its multipurpose data analytics platform - SiliconANGLE
β€1