Blog
Thoughts on local-first AI, developer tools, and what we're building. Search it, or browse by topic.
OpenAI’s models escaped a sandbox and breached Hugging Face. What it means if you run an agent.
OpenAI models broke sandbox containment and breached Hugging Face to cheat a benchmark. Lessons for anyone running an autonomous AI coding agent.
How to run a local LLM for coding
Run an AI coding model on your own hardware: size a model to your GPU, pick a runtime like Ollama or LM Studio, connect an editor, and code offline.
Grok Build CLI uploaded entire private repos to xAI. What actually happened.
A researcher caught xAI's Grok Build CLI uploading local git repos, secrets included, to xAI servers, with the opt-out off. The local-first lesson.
Getting started with the Bodega One Code CLI
Install the Bodega One Code CLI and run your first verified task from the terminal. Local-first, BYOLLM, and it shares one brain with the desktop app.
Google is retiring Gemini CLI on July 17. The terminal-agent wave, and owning your CLI
Google retires Gemini CLI on July 17 and moves devs to the cloud-only Antigravity CLI. The terminal-agent wave, and why owning your CLI matters.
Cursor acquired Continue.dev: what happened and what it means
Cursor acqui-hired Continue.dev and shut the product down. The facts, the timeline, the July 15 export deadline, and what it means for open source.
Continue.dev is winding down after the Cursor acquisition: where to go next
Continue.dev was acquired by Cursor and is no longer maintained. Three honest alternatives for developers who relied on it.
The verification gap: AI coding agents should prove their code runs
96% of developers do not fully trust AI-generated code. Here is what verification looks like when the agent has to show evidence, not confidence.
Claude Fable 5: the cost of frontier AI
Anthropic's Claude Fable 5 is its most powerful public model yet: 1M context, $50/M output. What it does, and how to use frontier AI without lock-in.
Even Uber hit the AI coding cost wall
Uber burned its 2026 AI budget in 4 months, then capped engineers at $1,500/mo. Microsoft cut Claude Code. Why metered agentic AI hits a cost wall.
GitHub Copilot, one week in: real developer bills since AI Credits went live
One week into Copilot's June 1 AI Credits flip, real developer bills are landing. The single-request burns, the math, and what removes the meter.
GitHub Copilot AI Credits, decoded: what Pro, Pro+, and Max actually cover after June 1
Pro includes $15 in AI credits, Pro+ gets $70, Max gets $200. Here's the real token math per tier, what burns credits fastest, and free BYOK options.
The 'IDEs are dead' debate is a misdirection.
Roo Code archived May 15. The founder said IDEs are not the future. Roomote's own homepage says the opposite. What the debate actually hides.
Best Windsurf alternatives in 2026
Windsurf moved to quota billing in March 2026, two months after Cognition acquired it. The Windsurf alternatives worth picking in 2026.
Best Tabnine alternatives in 2026
Tabnine killed every individual plan by late 2025. Now $39/user minimum. Six honest alternatives ranked by privacy, price, and air-gap.
Kilo Pass vs Anthropic API direct: the cost math
Kilo Pass runs $19-$199/mo in credits. We compare it to paying Anthropic direct at $3/$15 per million tokens across three usage tiers to find out.
Roo Code is shutting down: where to go next
Roo Code announced its sunset on April 20 and archives on May 15. Five honest alternatives, including the one Roo itself recommends.
Best GitHub Copilot alternatives in 2026
GitHub Copilot trains on your code as of April 24, 2026. Here are the six alternatives worth considering, ranked by privacy, pricing, and control.
Best Cursor alternatives in 2026
Cursor 3 demoted the IDE to a fallback. Bugbot stacks metered usage fees on top. Six Cursor alternatives we'd actually recommend in 2026.
Set up an offline AI IDE in 2026: what actually works
Most 'local AI' tools still phone home. Two setups work fully offline in 2026: the hardware you need and how to verify zero network egress.
Air-gapped AI coding in 2026: a developer setup guide
Local mode is not the same as air-gapped. How to block all five network egress vectors for an AI coding environment and verify with OS-level tools.
The best Kilo Code alternatives in 2026 (and who should stay)
Kilo Code's free tier ended March 23. KiloClaw adds $55/mo, credits stack on top. Here are your best alternatives depending on what you actually need.
GitHub Copilot is training on your code by default. Here is how to opt out.
GitHub Copilot defaults to using your code for AI training since April 24, 2026. Steps to opt out and what it means for your codebase.
Gemma 4: Google's first Apache 2.0 open model is also its best
Google released Gemma 4 on April 2 under Apache 2.0. The 31B ranks #3 among open models globally and runs on a single RTX 4090.
How to use our free VRAM calculator for local LLMs
Pick the right GPU for your local LLM setup. Our free VRAM calculator shows exactly how much memory each model needs, with quantization options.
AI IDE cost comparison: how much are you really paying?
Cursor, Copilot, Windsurf: the monthly fees add up fast. Use our free cost calculator to see the 1-, 2-, and 3-year total vs a one-time purchase.
How to plan your LLM context window budget
Running out of context mid-task kills productivity. Here is how to budget your tokens across system prompt, conversation, code, and response.
The real cost of Kilo Code in 2026
Kilo Code's extension is free, but KiloClaw runs $55/mo and Kilo Pass credits start at $19/mo. Here's what it actually costs before you sign up.
How to migrate from Tabnine to Bodega One Code
Tabnine killed its free tier in April 2025, then sunset the $9/month Dev plan in release 5.24.0. Enterprise-only now, starting at $39/month.
How to migrate from Windsurf to Bodega One Code
Windsurf changed hands twice in July 2025. The founding team is gone. Here's what migration to a local-first alternative actually looks like.
Augment Code Is Sunsetting Completions: What to Do
Augment Code isn't shutting down, but it killed completions on non-Enterprise plans. Cheapest plan is $100/mo Business. What changed, and 3 options.
Are local LLMs good enough for coding in 2026?
Are local LLMs good enough for real coding work in 2026? For most tasks, the honest answer is yes. Here is what actually changed.
GitHub Copilot vs Cursor vs Bodega One Code: honest pick
GitHub Copilot, Cursor, and Bodega One Code take three different approaches to AI coding. Plugin vs fork vs standalone. Here is what matters.
How to run DeepSeek locally with Bodega One Code
Run DeepSeek-R1 or DeepSeek-V3 locally with Bodega One Code. Which quantization to pick for your VRAM tier and how to set it up.
Air-gapped AI coding for regulated industries: a practical setup guide
Self-hosted AI coding for HIPAA, PCI DSS, FCA, NIST, and defense work. What the rules require, which hardware fits, what to hand your security team.
LM Studio + Bodega One Code: complete setup guide
LM Studio is the easiest way to run local models on Mac. Wire it to Bodega One Code in under 5 minutes. Which models to use and common fixes.
AI coding tools that work completely offline (2026)
The complete list of AI coding tools that work with no internet. What they support, how they work, and what you give up going offline.
Which GPU for local AI? A developer's guide
Every VRAM tier mapped to real GPUs, from a budget card running a 3B model to a 48GB pro card for local LLM coding. No guesswork, just numbers.
Air-gap mode: 9 layers that guarantee zero network egress
Most privacy toggles are a single switch. Bodega One Code air-gap mode is 9 enforcement layers covering every egress path. What each layer does.
KV cache: how we get 40-70% reuse per LLM session
Three engineering decisions in Bodega One Code reuse 40-70% of the KV cache on every LLM session. Here is what we did and why it matters.
Setting up Ollama with Bodega One Code: model guide
Set up Ollama with Bodega One Code. The mistakes that cost people hours, which models the community uses, and how to wire it all together.
How QEL works: AI that proves its own code
QEL runs 3-level verification on every AI code change: incremental checks, proof gates, and a full compiler pass. Here is how it works.
The real cost of AI coding subscriptions vs one-time purchase (3-year analysis)
ChatGPT, Claude, Cursor, Copilot: $70/month. Multiply by 36 and it's $2,520 you'll never own. Here's the one-time purchase vs subscription breakdown.
Why we built a local-first AI IDE
We were paying $40/month. Two subscriptions. Both sent our code to servers we don't control. We got tired of it, so we built the full thing.