[{"data":1,"prerenderedAt":191},["ShallowReactive",2],{"navigation":3,"index":14,"mdc--ptyor5-key":168},[4],{"title":5,"path":6,"stem":7,"children":8,"page":13},"Blog","\u002Fblog","3.blog",[9],{"title":10,"path":11,"stem":12},"Why Every Terminal AI Agent Codes for a Living","\u002Fblog\u002Fwhy-terminal","3.blog\u002F1.why-terminal",false,{"id":15,"title":16,"body":17,"cta":18,"description":32,"extension":33,"features":34,"hero":62,"meta":69,"navigation":13,"path":70,"sections":71,"seo":163,"stem":166,"__hash__":167},"index\u002F0.index.yml","Your [intelligence]{class=\"text-primary\"}, curated",null,{"title":19,"description":20,"links":21},"Ready to curate your expertise?","Install Working Mind. Teach it what you know. Verify the answers. When they are perfect, export. Fine-tune a domain champion.",[22,27],{"label":23,"to":24,"target":25,"trailingIcon":26},"Get Started","https:\u002F\u002Fdocs.workingmind.ai\u002Fgetting-started","_blank","i-lucide-arrow-right",{"label":28,"to":29,"target":25,"variant":30,"icon":31},"View on GitHub","https:\u002F\u002Fgithub.com\u002Fpawco\u002Fworking-brain","subtle","i-simple-icons-github","Working Mind is the local AI agent for domain experts. Instruct it, verify the answers, export your knowledge graph as training data, and fine-tune a domain champion that knows your field better than any frontier model.","yml",{"title":35,"description":36,"items":37},"Built for Privacy and Transparency","No telemetry. No accounts. No cloud. You install it, you run it, you own the data.",[38,42,46,50,54,58],{"title":39,"description":40,"icon":41},"Local First","Data stays on your disk. API keys go to LLM providers, not to us. Knowledge graph is a file you can inspect, backup, and delete.","i-lucide-shield",{"title":43,"description":44,"icon":45},"Fully Visible","Every tool call requires your approval. See exactly what the agent does, why, and with what arguments. No hidden actions.","i-lucide-eye",{"title":47,"description":48,"icon":49},"Your Keys, Your Providers","OpenAI, Anthropic, Google, Ollama, OpenRouter, or any OpenAI-compatible endpoint. You bring your own key. You choose your model.","i-lucide-key",{"title":51,"description":52,"icon":53},"MIT Licensed","Audit the code. Fork it. Build your own packs. No gatekeeper. The entire source is on GitHub.","i-lucide-scale",{"title":55,"description":56,"icon":57},"MCP Ecosystem","Connect to any MCP server for search, scraping, file operations, and more. The agent decides which tools to use. You approve.","i-lucide-plug",{"title":59,"description":60,"icon":61},"Works Offline with Ollama","No internet required when using local models. Your knowledge graph, your tools, your terminal. Completely self-contained.","i-lucide-wifi-off",{"links":63},[64,67],{"label":23,"icon":26,"trailing":65,"to":24,"target":25,"size":66},true,"xl",{"label":28,"icon":31,"size":66,"color":68,"variant":30,"to":29,"target":25},"neutral",{},"\u002F",[72,90,107,129,146],{"title":73,"description":74,"id":75,"orientation":76,"features":77},"How It Works","You teach. It learns. You verify. It compounds.","how-it-works","horizontal",[78,82,86],{"name":79,"description":80,"icon":81},"Instruct","Tell it what to learn. \"Remember that taaffeite has refractive index 1.719-1.730 and birefringence 0.004-0.005.\" It saves entities, relations, and observations to your local knowledge graph.","i-lucide-message-square",{"name":83,"description":84,"icon":85},"Verify","Ask a question. If the answer is wrong, fix the graph. Ask again. When the answer is perfect, the graph is right. You are the quality gate -- no pipeline can replace that.","i-lucide-check-circle",{"name":87,"description":88,"icon":89},"Compound","Next session, it remembers. Every conversation builds on the last. Cross-links emerge. Patterns surface. Your knowledge graph grows with every interaction.","i-lucide-brain",{"title":91,"description":92,"id":93,"orientation":76,"reverse":65,"features":94},"Persistent Knowledge Graph","Every fact, entity, and relationship is saved locally and recalled across sessions. This is not a chatbot with context window amnesia.","knowledge-graph",[95,99,103],{"name":96,"description":97,"icon":98},"Entity-Based Memory","People, projects, technologies, concepts, compounds, cases, statutes. Each gets its own node with observations attached.","i-lucide-database",{"name":100,"description":101,"icon":102},"Relations and Cross-Links","Taaffeite --confused-with--> Spinel. The agent connects entities when you mention them together. You see connections you forgot you made.","i-lucide-git-branch",{"name":104,"description":105,"icon":106},"Session Continuity","Close the app. Reopen it tomorrow. Ask \"What do I know about taaffeite?\" It remembers exactly what you taught it.","i-lucide-history",{"title":108,"description":109,"id":110,"orientation":111,"features":112},"Your Expertise, Exportable","You spent years learning your domain. You built a knowledge graph that produces perfect answers. Now export that expertise as training data and fine-tune a domain champion -- a small model that knows your field better than any frontier model.","your-intelligence","vertical",[113,117,121,125],{"name":114,"description":115,"icon":116},"Expert-Curated Data Is the Best Training Data","No synthetic pipeline, no crowdsourced annotations, no GPT-4-generated pairs can match knowledge you verified yourself. Every entity was explicitly created. Every observation was checked. Every relation was intentional. Shit in, shit out cannot be prevented by any pipeline -- but a domain expert who iterates until the graph is right eliminates the problem at the source.","i-lucide-graduation-cap",{"name":118,"description":119,"icon":120},"From Graph to Champion","When your graph produces perfect answers, export it. The .oexp format carries your full knowledge graph, taxonomy, and persona lenses. Convert it to instruction pairs. Fine-tune a 3.8B model in 30 minutes on a free GPU. Research shows 500 curated examples consistently outperform 5,000 auto-generated ones -- and break-even is ~100 labeled samples.","i-lucide-zap",{"name":122,"description":123,"icon":124},"Same Graph, Multiple Champions","Persona lenses let you export the same graph as different training data flavors. A researcher gets multi-hop reasoning pairs. An advisor gets strategic summary pairs. A student gets Socratic Q&A pairs. One knowledge graph, multiple domain champions -- each optimized for a different role.","i-lucide-users",{"name":126,"description":127,"icon":128},"You Own Both","The knowledge graph is on your disk. The fine-tuned model is on your machine. You can read the graph with a text editor, back it up with cp, delete it with rm. The model is a GGUF file you can run with Ollama. No cloud lock-in. No subscription. Your expertise, your models.","i-lucide-hard-drive",{"title":130,"description":131,"id":132,"orientation":76,"features":133},"The Pack System","Packs specialize Working Mind for your domain. A pack is a directory of declarative files -- a system prompt, MCP tools, skills, and personas -- that wires the agent for a specific field.","packs",[134,138,142],{"name":135,"description":136,"icon":137},"Starter Pack Included","Memory, web search, and scraping. Works immediately out of the box. No configuration needed.","i-lucide-package",{"name":139,"description":140,"icon":141},"Build Your Own","Copy the starter pack, edit the prompt, add tools, define skills. One directory, zero boilerplate. Your domain, your pack.","i-lucide-hammer",{"name":143,"description":144,"icon":145},"Domain Specialization That Works","Domain-specific packs measurably improve tool selection and task completion. Research shows brief structured reasoning before tool calls improves accuracy by up to 45%. Skills that decompose workflows reduce state-tracking errors. The architecture is sound.","i-lucide-check-circle-2",{"title":147,"description":148,"id":149,"orientation":111,"features":150},"Security First. Simplicity by Design. Not Scalable on Purpose.","Most AI products optimize for scale: more users, more data, more cloud. Working Mind optimizes for the opposite. One user. One machine. One knowledge graph. That constraint is the feature.","design-principles",[151,155,159],{"name":152,"description":153,"icon":154},"Security First","No remote servers. No web dashboard. No database to breach. MCP servers run as local child processes over stdio -- no HTTP, no sockets, no ports. Your data never leaves your machine except as API calls to the LLM provider you chose. Attack surface is the size of your terminal.","i-lucide-lock",{"name":156,"description":157,"icon":158},"Simplicity Is the Key","Knowledge graph is a JSONL file you can read with cat. Config is a JSON file you can edit with vim. Packs are directories of markdown. No database migrations. No schema registries. No service mesh. A working system you can understand end to end in an afternoon.","i-lucide-feather",{"name":160,"description":161,"icon":162},"Not Scalable by Design","No multi-tenancy. No concurrent writes. No distributed consensus. One user, one graph, one session at a time. This is not a limitation -- it is the architecture. Every scaling feature we did not build is a failure mode we do not have.","i-lucide-user",{"title":164,"description":165},"Working Mind - The Local AI Agent for Domain Experts","Working Mind is an open-source, local-first terminal AI agent with a knowledge graph that learns what you instruct. Export your expertise as training data. Fine-tune a domain champion. MIT licensed.","0.index","fr_95o4AjhKzdmB2Y8SHzI4MEAQLgEKnOHMlET5VY0w",{"data":169,"body":170},{},{"type":171,"children":172},"root",[173],{"type":174,"tag":175,"props":176,"children":177},"element","p",{},[178,181,189],{"type":179,"value":180},"text","Your ",{"type":174,"tag":182,"props":183,"children":186},"span",{"className":184},[185],"text-primary",[187],{"type":179,"value":188},"intelligence",{"type":179,"value":190},", curated",1778098907238]