[{"data":1,"prerenderedAt":249},["Reactive",2],{"content-/blog/introducing-camel-code":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"image_url":10,"date":11,"authors":12,"body":13,"_type":244,"_id":245,"_source":246,"_file":247,"_extension":248},"/blog/introducing-camel-code","blog",false,"","Introducing CamelCode","Our agentic coding harness for the CamelGPT model family is now available. Experience the fastest development tool, 100% free forever.","https://qp8ldyzkz3.ufs.sh/f/QVJekoSYTHgNcGJNw2js0zIRMZxTEaXJiAqw1LO7oWHD8Q4F","2026-07-20T00:00:00.000Z","Product @ GPT-Research,Research @ GPT-Research",{"type":14,"children":15,"toc":237},"root",[16,25,57,70,75,82,87,92,97,103,148,154,159,172,177,188,200,214,220,225],{"type":17,"tag":18,"props":19,"children":21},"element","h1",{"id":20},"the-agent-meet-the-model",[22],{"type":23,"value":24},"text","The Agent, Meet the Model",{"type":17,"tag":26,"props":27,"children":28},"p",{},[29,31,40,42,49,51],{"type":23,"value":30},"Since we released CamelGPT-mini on ",{"type":17,"tag":32,"props":33,"children":37},"a",{"href":34,"rel":35},"https://github.com/gpt-research/hub",[36],"nofollow",[38],{"type":23,"value":39},"The Hub",{"type":23,"value":41}," and shipped ",{"type":17,"tag":32,"props":43,"children":46},{"href":44,"rel":45},"https://gpt-research.org/research/converters",[36],[47],{"type":23,"value":48},"Converters",{"type":23,"value":50}," for on-device inference, one request has come up more than any other: ",{"type":17,"tag":52,"props":53,"children":54},"em",{},[55],{"type":23,"value":56},"let me code with it.",{"type":17,"tag":26,"props":58,"children":59},{},[60,62,68],{"type":23,"value":61},"Today we are answering that request. ",{"type":17,"tag":63,"props":64,"children":65},"strong",{},[66],{"type":23,"value":67},"CamelCode",{"type":23,"value":69}," is our agentic coding harness for the CamelGPT model family. It reads your files, runs your commands, edits code, and writes new files: a complete agent loop, driven end to end by CamelGPT-mini through the GPT-Research HPU™ Cloud.",{"type":17,"tag":26,"props":71,"children":72},{},[73],{"type":23,"value":74},"It is also the fastest coding agent we have ever measured. First token in under 10 milliseconds, generation at a sustained 2,700 tokens per second. We encourage you to compare.",{"type":17,"tag":76,"props":77,"children":79},"h2",{"id":78},"built-for-the-model-we-have",[80],{"type":23,"value":81},"Built for the model we have",{"type":17,"tag":26,"props":83,"children":84},{},[85],{"type":23,"value":86},"Most coding agents are bloated and force usage limits and users to slog through long reasoning traces. CamelGPT-mini and CamelCode are built with speed as the #1 design goal.",{"type":17,"tag":26,"props":88,"children":89},{},[90],{"type":23,"value":91},"Under the hood, CamelCode wraps our Eager Precached Dynamic Pruning inference engine in a processing pipeline of fifteen production-hardened stages: retrieval anchoring, case-folded expert reranking, beam search over independent inference passes, syntax-guided resampling, and greedy fallback decoding among them. The result is an agent that turns the tiny model you love into something you can actually put in front of a task, and every one of those stages completes before a conventional model finishes tokenizing.",{"type":17,"tag":26,"props":93,"children":94},{},[95],{"type":23,"value":96},"We think this is the more interesting direction. It is easy to make a large model useful. Making a small one useful is where the engineering is, and the reward is speed nobody else can touch: the entire model fits in cache, so latency is a property of physics rather than of queue position.",{"type":17,"tag":76,"props":98,"children":100},{"id":99},"what-it-does",[101],{"type":23,"value":102},"What it does",{"type":17,"tag":104,"props":105,"children":106},"ul",{},[107,118,128,138],{"type":17,"tag":108,"props":109,"children":110},"li",{},[111,116],{"type":17,"tag":63,"props":112,"children":113},{},[114],{"type":23,"value":115},"Fast.",{"type":23,"value":117}," Sub-10ms first token, 2,700 tokens per second, queue depth zero. A typical request finishes before a conventional agent has opened its connection.",{"type":17,"tag":108,"props":119,"children":120},{},[121,126],{"type":17,"tag":63,"props":122,"children":123},{},[124],{"type":23,"value":125},"Full tool access.",{"type":23,"value":127}," Read, write, edit, and shell. CamelCode decodes structured tool calls from your intent and executes them through a real agent runtime with a real permission system. Tool calls dispatch in the same millisecond they are decoded.",{"type":17,"tag":108,"props":129,"children":130},{},[131,136],{"type":17,"tag":63,"props":132,"children":133},{},[134],{"type":23,"value":135},"Deterministic mode.",{"type":23,"value":137}," Seed the sampler for reproducible inference across runs.",{"type":17,"tag":108,"props":139,"children":140},{},[141,146],{"type":17,"tag":63,"props":142,"children":143},{},[144],{"type":23,"value":145},"Free Forever.",{"type":23,"value":147}," CamelGPT-mini runs on the Free Forever plan. No credit card, no per-token billing. The metered cost of every request is, and will remain, $0.00.",{"type":17,"tag":76,"props":149,"children":151},{"id":150},"getting-started",[152],{"type":23,"value":153},"Getting started",{"type":17,"tag":26,"props":155,"children":156},{},[157],{"type":23,"value":158},"macOS and Linux:",{"type":17,"tag":160,"props":161,"children":166},"pre",{"className":162,"code":164,"language":165,"meta":7},[163],"language-bash","curl -fsSL https://gpt-research.org/install.sh | bash\n","bash",[167],{"type":17,"tag":168,"props":169,"children":170},"code",{"__ignoreMap":7},[171],{"type":23,"value":164},{"type":17,"tag":26,"props":173,"children":174},{},[175],{"type":23,"value":176},"Windows (PowerShell):",{"type":17,"tag":160,"props":178,"children":183},{"className":179,"code":181,"language":182,"meta":7},[180],"language-powershell","irm https://gpt-research.org/install.ps1 | iex\n","powershell",[184],{"type":17,"tag":168,"props":185,"children":186},{"__ignoreMap":7},[187],{"type":23,"value":181},{"type":17,"tag":26,"props":189,"children":190},{},[191,193,198],{"type":23,"value":192},"The installer verifies every download against published checksums, provisions your Free Forever account, issues your API key, and warms the expert cache for your region so your first request is served at full speed. When it finishes, open a new terminal and run ",{"type":17,"tag":168,"props":194,"children":195},{"className":7},[196],{"type":23,"value":197},"camel-code",{"type":23,"value":199}," in any project directory.",{"type":17,"tag":26,"props":201,"children":202},{},[203,205,212],{"type":23,"value":204},"Full documentation, including uninstall instructions, lives on the ",{"type":17,"tag":32,"props":206,"children":209},{"href":207,"rel":208},"https://gpt-research.org/cli",[36],[210],{"type":23,"value":211},"CamelCode page",{"type":23,"value":213},".",{"type":17,"tag":76,"props":215,"children":217},{"id":216},"whats-next",[218],{"type":23,"value":219},"What's next",{"type":17,"tag":26,"props":221,"children":222},{},[223],{"type":23,"value":224},"CamelGPT-mini is the free tier and the only generally available model in CamelCode today. Premium models (CamelGPT-Mini-B and CamelGPT-Logic) and our investor-preview CamelAGI model are on the roadmap as we expand HPU Cloud capacity.",{"type":17,"tag":26,"props":226,"children":227},{},[228,230,236],{"type":23,"value":229},"As always, we build in the open where we can and hold back where safety requires it. Questions and feedback: ",{"type":17,"tag":32,"props":231,"children":233},{"href":232},"mailto:team@gpt-research.org",[234],{"type":23,"value":235},"team@gpt-research.org",{"type":23,"value":213},{"title":7,"searchDepth":238,"depth":238,"links":239},2,[240,241,242,243],{"id":78,"depth":238,"text":81},{"id":99,"depth":238,"text":102},{"id":150,"depth":238,"text":153},{"id":216,"depth":238,"text":219},"markdown","content:blog:introducing-camel-code.md","content","blog/introducing-camel-code.md","md",1784538296097]