installation0

<p align="center"> <img src="installation0.jpg" width="100%" alt="INSTALLATION0 — local AI toolkit for Claude Code & GitHub Copilot"> </p> # INSTALLATION0 **The root repository for a local-AI toolkit** — a suite of modules for Claude Code and GitHub Copilot, installed as one and unified by a single shared server instance. INSTALLATION0 is the home for all of these tools. One `node install.js` sets up a set of modules that plug into Claude Code and Copilot: | Module | What it is | |---|---| | **bridge** | the flagship — Copilot Chat → local `claude` (BYOM). Detailed below. | | **engram** | cross-project agent memory + reference dictionary (MCP) | | **code-memory** | graph + full-text index of your codebase (MCP) | | **filesystem** (`fsp`) | token-cheap filesystem proxy (CLI + `Bash` hook) | | **local-subagent** | orchestrate local-LLM subagents from Claude (MCP) | They're tied together by a **primary shared server instance** — the `server` module — the always-on backbone that owns cross-module **analytics, telemetry, and the data spine**: a DuckDB events store plus per-bridge SQLite drill-downs, unified under one CLI / HTTP / MCP surface. Everything the modules do reports into this one place. Everything runs **on your machine**. No cloud egress beyond the model calls you already make. --- ## Install ```sh git clone https://github.com/Installation-0/installation0.git cd installation0 node install.js ``` Or one line — checks git/Node, clones, then runs the installer: ```sh curl -fsSL https://raw.githubusercontent.com/Installation-0/installation0/main/bootstrap.sh | bash ``` `node install.js` stands up the bridge as an HTTP service on `127.0.0.1:8400` and registers the supporting MCP servers. Then add it to VS Code as a BYOM endpoint and pick **installation-0** in the Copilot Chat model dropdown. Flags: `--lan`, `--autostart`, `--hooks`, `--only=a,b`. Full BYOM wiring, requirements, autostart, and uninstall are in **[INSTALL.md](INSTALL.md)**. > After install, fully quit and reopen VS Code / Claude Code so it reloads > `~/.claude.json`, then `/mcp` to confirm the servers are live. --- ## The bridge — Copilot → local `claude` > **An agnostic VS Code Copilot Chat ↔ Claude Code bridge.** > Use Copilot Chat — its UI, its native tools, its approvals — powered by > **Claude Code**, through a Bring-Your-Own-Model server bridge to your local > **subscription** `claude`. No API key, no per-token cost. Point GitHub Copilot Chat at `installation-0` in the model picker and every message you send is answered by the local `claude` CLI you already pay for — streaming back into Copilot's own chat panel, calling Copilot's own tools, under Copilot's own approvals. The bridge is invisible; only the engine changes. <p align="center"> <img src="byom.png" width="420" alt="GitHub Copilot Chat with installation-0 selected as the model"> </p> <p align="center">_{<i>Select <b>installation-0</b> as the model — Copilot Chat is now driving Claude Code.</i>}</p> Because it speaks the plain OpenAI chat API, it's **agnostic** — Copilot Chat drives it here, but any OpenAI-compatible client can. Added through VS Code's **Manage Models** flow as a local BYOM endpoint, it runs with **no Copilot plan and no GitHub sign-in** — local models are free in VS Code. Only Copilot-service features (semantic search, inline suggestions, embeddings) still need a plan; chat and agent mode with `installation-0` do not. [(VS Code docs.)](https://code.visualstudio.com/docs/agent-customization/language-models#_frequently-asked-questions) | Capability | What you get | |---|---| | **BYOM model** | `installation-0` appears in Copilot's model picker → routes to local `claude` | | **No cost** | your **subscription** CLI — no API key, no per-token billing, and as a local BYOM endpoint it needs **no Copilot plan and no GitHub sign-in** | | **Native tools, both ways** | claude calls VS Code's own edit / terminal / diagnostics / search tools | | **Streaming** | real-time tokens *and* tool-call deltas inside Copilot Chat | | **Persistent sessions** | one resumable `claude` session per conversation — context survives every turn | | **`/model`** | switch `opus` / `sonnet` / `haiku` (or a full `claude-…` id) mid-chat | | **`/effort`** | `low` → `max` reasoning effort on demand | | **`/perms`** | `bypass` / `acceptEdits` / `plan` / `default` permission mode | | **`/compact`** | compact now — plus an automatic compaction governor | | **Local & private** | loopback by default; `--lan` to share, `--autostart` to run as an OS service | ### Copilot Chat, Claude Code engine You keep everything you like about Copilot Chat — the panel, agent mode, checkpoints, the approvals UI — and swap the brain underneath for Claude Code. The bridge keeps **one persistent `claude` stream-json session per conversation**, resumed across turns by a conversation fingerprint, so context and session state persist instead of starting cold on every message. ### Native VS Code tools, in both directions claude doesn't just talk — it **acts in your editor**. Through an MCP proxy, a claude tool call becomes an OpenAI `tool_calls` delta that Copilot executes with its *own* native tools (edit files, run the terminal, read diagnostics, list changed files, manage the todo list, search the workspace) under its own approval flow — and the result streams back to claude. <p align="center"> <img src="vscode_tools.png" width="300" alt="VS Code tools claude can call through the bridge"> </p> <p align="center">_{<i>The VS Code tools claude can invoke through the bridge.</i>}</p> ### Drive claude from the chat box Copilot Chat has no UI for reasoning effort, model, or permissions — so the bridge intercepts a small command vocabulary and answers inline. Each one restarts the session **with your context preserved**, so nothing is lost: | Command | Does | |---|---| | `/model <opus\|sonnet\|haiku>` | switch claude's model (or a full `claude-…` id) | | `/effort <low\|medium\|high\|xhigh\|max>` | set reasoning effort | | `/perms <bypass\|acceptEdits\|plan\|default>` | set permission mode | | `/compact` | compact this conversation now | | `/help` | list the commands | <p align="center"> <img src="effort_support.png" width="760" alt="/effort command handled by the bridge"><br> <img src="model_selection.png" width="760" alt="/model command handled by the bridge"> </p> <p align="center">_{<i><code>/effort</code> and <code>/model</code>, handled inline without leaving Copilot Chat.</i>}</p> ### Never hit the wall A **compaction governor** watches the context window: it warns as the window fills, and at the wall it halts new input, **queues** your messages, compacts the conversation, and **resumes automatically** — so a long session continues instead of dead-ending. ### Permissions & approvals `/perms` threads the permission mode straight to claude — `bypass` for hands-free, `plan` for read-only, `acceptEdits` to auto-accept edits while gating shell commands. The Copilot tools claude calls run under **Copilot's own auto-approve**, so you control the editor side from the UI you already use. ### Local, private, always-on The bridge binds to `127.0.0.1:8400` by default. `--lan` exposes it on a network you trust; `--autostart` installs it as an OS service (launchd / systemd) that survives reboots. No cloud egress beyond the `claude` calls you already make. Full reference: **[bridge/README.md](bridge/README.md)**. --- ## The rest of the suite The bridge is the headline, but the modules it ships alongside are what make the **local AI nexus** — Claude Code and the bridge draw on all of them: - **engram** — cross-project **agent memory** + reference dictionary, graph-connected (SurrealDB). MCP server + skill. - **code-memory** — a **graph + full-text index of your codebase**: symbol lookup, call-path tracing, impact analysis, architecture overviews, LSP pass-through. MCP server + skill. - **filesystem** (`fsp`) — a **token-cheap filesystem proxy**: file ops go through a `Bash` hook so raw bytes stay off the agent's context window. - **local-subagent** — orchestrate many **local-LLM subagents** (LM Studio / any OpenAI-compatible endpoint) from Claude — text-only workers, zero cloud egress. MCP server + skill. - **server** — the **shared analytics & data spine** introduced up top: the DuckDB + SQLite backbone every module reports into. Query it to see what your agents have been doing. See **[server/README.md](server/README.md)**. --- ## Requirements - **Node ≥ 20** - **pnpm** — `npm i -g pnpm` - **A C toolchain** (cc/clang + make) — some components compile native modules (better-sqlite3, lmdb) on install. macOS: `xcode-select --install`; Debian/Ubuntu: `apt install build-essential`. The installer hard-checks all three up front and stops with a fix if any are missing. ## Security Defaults are safe: the bridge binds to **loopback (`127.0.0.1`)** with local-AI features **off**. Its admin/mutation endpoints have **no auth** — only expose it on a LAN (`--lan`) on a network you trust. See **[SECURITY.md](SECURITY.md)**. ## Uninstall ```sh node install.js --uninstall # reverse the registration (MCP servers, hooks, config) node uninstall.js # the above + remove node_modules / pnpm artifacts (keeps your data) ``` ## License [PolyForm Noncommercial 1.0.0](LICENSE) — source-available, **free for any noncommercial use** (personal, research, education, non-profit). Commercial use is not granted, and the copyright/attribution notice must be preserved. For a commercial license, contact the author.