T3N

solana:DU7Lhe2QtcRBJdor7KbpjfUDmLyA2kdhR49Li3czpump ELI5 Imagine a workshop of small helpers, each very good at one thing. Oracle7 reads your question, works out what you want and what you gave it, and the parts that fit pass it along until it’s answered. Every answer shows its working: which parts it used and which records it read. If it doesn’t know, it says what’s missing instead of guessing. That’s a hole, and holes are where it grows: a new part is built, and kept only if it passes its test. No language model inside. It runs on an ordinary computer. Free and open: huggingface.co/psiloceyeben

How to use Oracle7 with Claude, ChatGPT, Hermes and the rest Your assistant is good at breadth. It can attempt almost anything, in fluent words. But it predicts its answers, so it can’t always tell you whether a particular one is right. Oracle7 works the other way. It has no language model inside. It works answers out from checked parts and named records, shows its working, and when it doesn’t know, it says so and names what’s missing. Its error rate is measured in sealed tests, run once and published even when they fail. Put the two together and each covers the other’s weak side. Your assistant keeps the conversation and the breadth. Oracle7 takes the repeated work and checks the rest. Two ways to use them together 1. Handoff. Your assistant passes Oracle7 the questions it keeps getting. Each one comes back answered with its evidence, or declined with the reason. Your assistant only has to deal with the declines. In one recorded session, Claude handed over 26 tasks. Oracle7 answered 23 with their evidence and declined 3, in 11.9 seconds. 2. Check. Your assistant writes an answer, and Oracle7 works the same question out on its own. It returns one of three verdicts: confirmed, contradicted (with both answers shown), or can’t check. It never confirms what it can’t check. How it plugs in Oracle7 runs as an MCP server (Model Context Protocol) on your own machine. Your assistant’s harness starts it and calls it like any other tool. It offers six tools: ask: one question, answered with its evidence or refused with a reason check: your assistant’s answers, each confirmed, contradicted or marked can’t check handoff: a batch of repeated tasks verify: checks the install against its published release hashes operations: lists everything this install can do, with input schemas invoke: calls any one formula part or command directly Oracle7 prints the setup for your harness: python oracle7.py mcp --config claude # Claude Code python oracle7.py mcp --config codex # OpenAI Codex python oracle7.py mcp --config hermes # Nous Research Hermes Agent python oracle7.py mcp --config cursor # Cursor python oracle7.py mcp --config gemini # Gemini CLI For Claude Code it’s one line, claude mcp add oracle7 -- python oracle7.py mcp, and then you just ask Claude to check something with Oracle7. When your recurring work isn’t covered yet Build a specialist for it. You don’t need to write code, and nothing is trained: From worked examples: python oracle7.py compose --file examples.txt, with 8 or more lines like 1.050, 1.010 -> 5.25. Half the examples write the program and half examine it. It’s admitted only if it gets the examined half right. From example questions: python oracle7.py build "q1" "q2" "q3", with 3 to 6 questions. The swarm works out what they need and names the gap. Where the gap is a table of named things, it builds the table from Wikidata, with every row cited. Either way, the new specialist then sits one exam on questions held back from it. After that, the job runs as a program: the same answer every time, in under a second, with its working. And every new specialist joins the network, so it can combine with the others. Why Use Oracle7 Answers you can check. The working and the records travel with each one. An error rate that’s measured, not claimed. A second opinion that fails differently, because it computes and looks things up instead of predicting. Repeated work stops costing a model call every time. It runs on an ordinary computer, with no graphics card, so your questions stay on your machine. Free and open: huggingface.co/psiloceyeben/… Compatible harnesses and models Oracle7 works with any harness that can start a local MCP server and call its tools. The model comes from the harness, so any model your harness runs can use Oracle7, provided the model handles tool calls. Setup printed by oracle7.py mcp --config: Claude Code (Anthropic) Codex (OpenAI) Hermes Agent (Nous Research) Cursor Gemini CLI (Google) Other MCP clients use the same standard and launch command (python oracle7.py mcp), but we haven’t tested each of these ourselves: Claude Desktop Goose Cline Zed OpenCode Windsurf LM Studio ChatGPT: the ChatGPT app connects only to remote MCP servers, and Oracle7 ships as a local one. The OpenAI route today is Codex. Models, through those harnesses: Claude (Opus, Sonnet, Haiku) GPT and OpenAI’s open-weight gpt-os Gemin Hermes (Nous Research) Kimi K2 (Moonshot) Qwen (Alibaba) DeepSeek GLM (Zhipu) Llama (Meta) Mistral The open models above run locally or through a provider such as OpenRouter, from a harness like Hermes Agent, Goose, Cline or OpenCode.