Invincible sits between your AI clients and your providers — routing chat across a tiered failover chain, executing agent tools under per-client approval, and remembering context across sessions.
$ pip install invincible && invincible start ✓ Gateway listening on :8000 ✓ Providers loaded: 5 (tiered failover armed) ✓ MCP tool server ready at /mcp → Point any OpenAI-compatible client at http://localhost:8000/v1
From install to AI-ready with zero cloud lock-in. Register an account, connect your keys, and route — no extra setup.
One pip install, one command. Works on macOS, Linux, Windows.
pip install invincible
invincible start
Create an account and the dashboard opens up — bring your own provider keys, they stay yours.
→ /register in your browser
Your keys, your data — isolated per account.
Point any OpenAI-compatible or Anthropic-native client at the gateway. Approve MCP clients one by one from the dashboard.
base_url: http://localhost:8000/v1
api_key: <your key>
Every surface carries its own realm — a chat key can only spend credits, a tool client can only run what you approved, and the dashboard answers only to your login.
A rate-limited free tier no longer kills an agent's workflow: requests fail over down the chain automatically. MCP tool clients pair through an owner-consent OAuth flow — every client is individually revocable. Chat credentials can never touch management surfaces.
One process, one database — no external services.
Requests cascade across providers on 429s and errors — pick the cheapest tier that's alive, automatically.
read_file, write_file, execute_bash — behind per-client OAuth tokens and a denylist + approval safety net.
Task state, checkpoints, and failover history persist per session — agents resume where they left off.
Each user's memories inject into their own prompts only — budget-capped, provenance-tracked.
Add, disable, and test providers at runtime; route auto, pinned, or chained — no restarts.
Per-provider attempts, failovers, token spend, and latency — visible in the dashboard, recorded per run.