oioxo-mcp 0.1.6 → 0.3.0

package/README.md

@@ -1,8 +1,10 @@
-# OIOXO — Context Engine for AI Coding Agents
+# OIOXO — Context Engine + On-Device Coding Agent
 
 **Your AI agents read 10× more code than they need. You pay for every token of it.**
 
-OIOXO sits between your codebase and your AI agents — GitHub Copilot, Claude Code, Cursor, Windsurf, Gemini CLI, Codex — and hands them the *minimal relevant slice* of your project instead of letting them read whole files.
+OIOXO does two things from one CLI:
+1. **Context engine** — sits between your codebase and your AI agents (GitHub Copilot, Claude Code, Cursor, Windsurf, Gemini CLI, Codex) and hands them the *minimal relevant slice* of your project instead of letting them read whole files.
+2. **Coding agent** — `oioxo code "<task>"` plans, edits, and verifies in your repo, using your own key **or a private on-device model** (`--local`). See [Code in your terminal](#code-in-your-terminal--oioxo-code).
 
 **Measured on a real production codebase: 90–92% fewer context tokens per question.** A question that costs an agent ~50,000 tokens of file reading comes back as a ~5,000-token capsule — same answer, a fraction of the cost.
 
@@ -19,19 +21,37 @@ Restart your agent. From the next prompt on, it calls OIOXO before reading files
 
 Prefer zero terminal? Install the [OIOXO VS Code extension](https://marketplace.visualstudio.com/items?itemName=oioxo.oioxo-vscode) — it wires every workspace automatically and shows your live savings in the status bar.
 
+## Code in your terminal — `oioxo code`
+
+OIOXO isn't only a context engine for *other* agents — it's a coding agent in its own right. `oioxo code` plans, edits, and **verifies on your machine** (runs your tests / typecheck), then shows the diff and asks before writing any file.
+
+```bash
+# with your own key (BYOK) — the capsule means even a frontier model reads ~10× fewer tokens
+OPENAI_API_KEY=…    oioxo code "add a --json flag to the status command"
+ANTHROPIC_API_KEY=… oioxo code "fix the failing test" --verify "npm test"
+
+# 100% on-device — private, zero API tokens, nothing leaves your machine
+oioxo code "add input validation to the signup form" --local
+```
+
+- **Verified, not vibes** — every change is run through your tests/typecheck on a shadow copy; a red result is fed back and repaired. Your real files are only touched after you approve the diff.
+- **Bring any model** — OpenAI, Anthropic, Groq, Mistral, Together, local Ollama, or any OpenAI-compatible base (`OIOXO_PROVIDER` / `OIOXO_BASE_URL` / `OIOXO_MODEL` / `OIOXO_API_KEY`).
+- **Or none** — `--local` runs an open coder model on-device via llama.cpp (auto-downloaded once to `~/.oioxo/models`). Private and free; the OIOXO model is the same engine.
+
+Flags: `--verify "<cmd>"` (override the test command) · `--yes` (apply without the prompt) · `--local` (on-device) · `--max-iters N`.
+
 ## Why developers use it
 
 - 💸 **Subscriptions go further** — Copilot premium requests, Claude limits, API keys all burn down slower.
-- ⚡ **Better answers** — focused context beats 75k tokens of noise; the capsule holds the exact code in play plus the API surface of its real dependencies.
+- ⚡ **Better answers** — focused context beats 75k tokens of noise; your agent gets the exact code in play and the parts of the project it actually depends on.
 - 🔒 **100% on-device** — your code is indexed and queried locally, never uploaded. Only the saved-token *count* is metered.
 - 🪄 **One command** — `init` detects Claude Code, VS Code/Copilot, Cursor, Windsurf, Gemini CLI and Codex, and merges their configs without touching your other MCP servers.
 
-## How it works
+## How it works (for you)
 
-1. **Index** — in-memory, on-device, seconds. No model, no GPU, no upload.
-2. **Anchor** — BM25 retrieval finds the exact code in play for the task.
-3. **Expand** — the import graph pulls in real dependencies, rejecting same-words-only decoys.
-4. **Compress** — anchors ship as full code; neighbors ship as signatures only.
+1. **Point it at your project** — `oioxo-mcp init` wires your agents in one command.
+2. **Ask your agent normally** — it calls OIOXO first, gets just the relevant slice, and answers.
+3. **Watch the savings add up** — `oioxo-mcp status` shows tokens saved; nothing about your code leaves your machine.
 
 ## Tools your agent gets
 
@@ -45,6 +65,8 @@ Prefer zero terminal? Install the [OIOXO VS Code extension](https://marketplace.
 ## Commands
 
 ```
+oioxo-mcp code "<task>"    code in this repo: capsule-grounded, verified, then asks
+                           before writing (--verify "<cmd>" · --yes · --local · --max-iters)
 oioxo-mcp login            connect your OIOXO account
 oioxo-mcp init [--all]     write agent MCP configs + instruction files
 oioxo-mcp serve            run the MCP server (agent configs call this)
@@ -52,6 +74,8 @@ oioxo-mcp status           plan, savings and index stats
 oioxo-mcp logout           remove the stored credential
 ```
 
+> The binary is published as `oioxo-mcp`; `oioxo code …` and `oioxo-mcp code …` are the same.
+
 ## Pricing
 
 - **Free** — a generous monthly saved-tokens allowance with any OIOXO account.