@vernikr/gemini-mcp 0.8.0 → 0.8.2

package/README.md

@@ -1,19 +1,88 @@
 # @vernikr/gemini-mcp
 
-One-command Gemini **free-tier** MCP gateway (Python core: `gemini-mcp-gateway` on PyPI,
-bootstrapped via [uv](https://docs.astral.sh/uv/) — no system Python needed). Works
-**without any server**: default `direct` mode talks to Google AI Studio straight from
-your machine; `router` mode joins a fleet gateway with a shared quota ledger.
+Gemini **free-tier** MCP gateway that needs **no server**: your MCP client (Freebuff,
+Claude Desktop, Cursor, …) talks to Google AI Studio through a quota-aware router —
+cascading Flash chain, per key×model RPD/RPM/TPM budgeting, minute-limit waiting,
+massive-body chunked analysis with consolidation. Python core `gemini-mcp-gateway`
+(PyPI), bootstrapped via [uv](https://docs.astral.sh/uv/) — no system Python needed.
+
+## Get API keys (free)
+
+<https://aistudio.google.com/app/apikey> — several keys = several times the quota
+(20 requests/day per key×model).
+
+## Wiring — 2 options
+
+### A · Wizard (recommended): `setup`, then paste one JSON block
 
 ```bash
-pnpm dlx @vernikr/gemini-mcp setup     # wizard: AI Studio keys → ~/.config/gemini-mcp
-                                       #         → live key probe → client snippets
-pnpm dlx @vernikr/gemini-mcp doctor    # self-check with fix commands
-pnpm dlx @vernikr/gemini-mcp quota     # free-tier budget table
-pnpm dlx @vernikr/gemini-mcp uninstall --yes   # remove keys+config
+pnpm dlx @vernikr/gemini-mcp setup
 ```
 
-Features: quota-aware routing across the Flash chain (two routes: `optimal` /
-`best_available`), massive-body chunked analysis with consolidation, file ingest with
-server-side RTF/DOCX→Markdown extraction, async tickets for long jobs, structured
-agent-facing errors. Repo: github.com/vernikr/gemini-mcp
+What it does — and what it does NOT:
+
+- asks one question (simple/direct vs fleet/router; pick **1**);
+- takes keys through a **masked** prompt and stores them in `~/.config/gemini-mcp/.env`
+  (mode `0600` — the only place keys live), plus a minimal `config.yaml`;
+- probes keys live (listModels — zero quota spent) and prints fingerprints, never keys;
+- **prints** the ready `mcpServers` JSON for your client — it does **not** edit any
+  client config itself (formats differ per client; you paste the block once).
+
+After the wizard the client JSON needs no env/keys — the gateway auto-discovers
+`~/.config/gemini-mcp/`:
+
+```json
+{
+  "mcpServers": {
+    "gemini": { "command": "npx", "args": ["-y", "@vernikr/gemini-mcp", "stdio"] }
+  }
+}
+```
+
+### B · JSON-only (skip the wizard entirely)
+
+Is pasting JSON into `mcp.json` enough? **Yes — if the keys travel in its `env`**
+(without keys anywhere the gateway fails fast by design):
+
+```json
+{
+  "mcpServers": {
+    "gemini": {
+      "command": "npx",
+      "args": ["-y", "@vernikr/gemini-mcp", "stdio"],
+      "env": { "GEMINI_API_KEYS": "AIza…key1,AIza…key2" }
+    }
+  }
+}
+```
+
+Backend auto-resolves to `direct`; logs and the local usage ledger (`state.db` —
+quota counters survive restarts) go to `~/.config/gemini-mcp/`. Trade-off vs option A:
+keys sit in the client config file instead of a `0600` `.env`.
+
+Where to paste (restart the client fully afterwards): Freebuff `~/.agents/mcp.json` ·
+Claude Desktop `~/Library/Application Support/Claude/claude_desktop_config.json` ·
+Cursor `~/.cursor/mcp.json`. First launch pays a one-time uv resolve (~10–20 s).
+
+## Commands
+
+```bash
+pnpm dlx @vernikr/gemini-mcp doctor          # keys, chain, limits, live probe, ledger
+pnpm dlx @vernikr/gemini-mcp quota           # per key×model budget table (live)
+pnpm dlx @vernikr/gemini-mcp calls -n 20     # local audit tail
+pnpm dlx @vernikr/gemini-mcp uninstall --yes # remove keys, config and ledger
+```
+
+All commands accept `--json` (agent-facing output). MCP tools: `analyze_large`,
+`gemini_generate`, `analyze_file` (server-side RTF/DOCX extraction), `analyze_result`,
+`quota_status`, `router_logs`. Every compute answer starts with a one-line `report`
+(`backend=direct · route=… · model=… · req=N · …s · clean run`).
+
+## How the shim works
+
+`bin/gemini-mcp.js` is a ~30-line bootstrap: it runs the version-pinned Python core
+(`uvx --from gemini-mcp-gateway==<same version> gemini-mcp …`), self-installing uv
+through `npx -y @astral-sh/uv` when needed. Zero logic in JS; npm ↔ PyPI versions are
+lockstep. Fleet mode (shared quota ledger on your own server) and full docs:
+[github.com/vernikr/gemini-mcp](https://github.com/vernikr/gemini-mcp) —
+`docs/06-consumer.md`, `docs/08-dual-mode.md`.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@vernikr/gemini-mcp",
-  "version": "0.8.0",
+  "version": "0.8.2",
   "description": "One-command Gemini free-tier MCP gateway: quota-aware model routing, massive-body chunked analysis, file ingest (RTF/DOCX→Markdown), async tickets. Thin bootstrap for the Python core gemini-mcp-gateway.",
   "bin": {
     "gemini-mcp": "bin/gemini-mcp.js"