Create an account
Register with email and password, or sign in with GitHub when your deployment enables it. Your account owns its keys, providers, sessions, and state.
Unlocks chat + dashboard + memory
/registerRemote-first · Per-user BYOK · Owner-scoped state
Connect your own provider credentials, route AI clients through one stable endpoint, and keep sessions, memories, task state, and checkpoints scoped to your account. Add OAuth-protected MCP tools and route confirmed commands to your paired machine when you need them.
pip install invincible-ai invincible agent # pairs your machine — no database, no .envOne command pairs your PC with invincible-ai.me. Your account, provider keys, and memory live on the hosted service.
A free account unlocks chat instantly. Pair the agent to give a remote AI hands on your machine. Connect provider keys to power every model route.
$ pip install invincible-ai $ invincible agent ✓ Paired with invincible-ai.me base https://invincible-ai.me/v1 key inv_… ✓ shown once · revocable chat OpenAI Chat · Anthropic Messages · Responses (Codex) tools https://invincible-ai.me/mcp ✓ OAuth
Get started
Each step unlocks a different part of Invincible. You do not need all three to begin — chat works before any machine is paired. Hosted at invincible-ai.me: no local database, environment file, or server process.
Register with email and password, or sign in with GitHub when your deployment enables it. Your account owns its keys, providers, sessions, and state.
Unlocks chat + dashboard + memory
/registerChoose a provider from the catalog or add any OpenAI-compatible Chat Completions endpoint. Each key is encrypted at rest and used only for your account's requests.
Unlocks every model route
dashboard → providersGenerate a per-user inv_ key from Account. The raw value is shown once, stored as a hash, and can be revoked immediately.
Unlocks /v1 chat from any client
dashboard → account → API keysPoint any OpenAI-compatible client or Codex at https://invincible-ai.me/v1, or Claude Code via ANTHROPIC_BASE_URL=https://invincible-ai.me. Then pair the agent so approved tool calls run on your PC.
Unlocks hands on your machine
pip install invincible-ai · invincible agentFour pieces, one system
Every piece stands on its own for the clients you already use. Together, they keep one stable endpoint, one memory, and one task history across every model you route through.
One stable endpoint for every model. Per-user BYOK failover across auto, pinned, or chain routing — a 429 or 5xx cools down and moves to your next credential instead of killing the run.
Needs: your own provider keys — there is no shared pool.
Agents that remember across tools and sessions. Scoped lexical retrieval over PostgreSQL blends full-text relevance, recency, kind, and confidence under one bounded budget. Say "remember this" and it is saved.
Needs: nothing extra — memories accrue as you chat.
Task state that survives provider switches. Versioned states with optimistic checks, immutable checkpoints, and one pre-failover snapshot per request when task state exists.
Needs: any chat client — state versions itself as you work.
OAuth-protected MCP tools with real hands on your machine. Ten tools across files, shell, task state, and memory — destructive verbs stage behind a single-use confirmation token, then run on your paired agent inside a home-relative sandbox.
Needs: invincible agent running on the machine you want to reach.
Ways to use Invincible
No machine needed for chat and memory. Pair a machine when you want a remote AI to act on your files. Open the dashboard when you want to see how every run moved.
Mint an inv_ key and point any OpenAI-compatible client at https://invincible-ai.me/v1. Sessions, memories, and continuity accrue immediately — no install, no pairing.
Run invincible agent on the machine you want to reach, then connect any MCP client to https://invincible-ai.me/mcp. Confirmed commands and writes execute on your PC — zero inbound ports.
Connect provider credentials and every route, failover, checkpoint, and token count shows up in the dashboard — sessions, task board, memory graph, and per-day usage by provider.
Explicit boundaries
Each surface uses its intended credential. Chat keys authorize /v1/*, OAuth tokens protect /mcp, and browser sessions protect account management.
Your keys. Your models. One endpoint.
Add a provider key on the dashboard and every model it can reach shows up in one place. Provider keys stay server-side — they never touch your client.
curl https://invincible-ai.me/v1/chat/completions \ -H "Authorization: Bearer inv_…" \ -H "X-Session-Id: my-conversation" \ -d '{"messages": [{"role": "user", "content": "Hello!"}]}'
History under the session id is prepended upstream; the reply is persisted back. A 429 cools down and fails over — the client sees one stable endpoint.
# Claude Code — Anthropic Messages ANTHROPIC_BASE_URL=https://invincible-ai.me # Codex — OpenAI Responses https://invincible-ai.me/v1/responses # model is a soft hint; the response # names the credential that served it
Both protocols translate into one internal message model, so an OpenAI client and Claude Code on the same session id see the same conversation.
{
"mcpServers": {
"invincible": {
"url": "https://invincible-ai.me/mcp"
}
}
}Approve the connector in the browser — OAuth 2.1 + PKCE mints the bearer token. Destructive calls return a confirmation token first; nothing runs until you approve it.
Continuity and control
Provider choice stays replaceable. Your account, projects, credentials, memory, and continuity state remain attached to Invincible.
Choose auto, pinned, or chain routing. Invincible tries the next available credential after rate limits, server errors, network failures, or disabled authentication — without claiming to optimize for price.
Task state is versioned with optimistic checks, immutable checkpoints pin important states, and one pre-failover snapshot captures the provider failure before routing continues — only when task state exists.
PostgreSQL stores memories with layer, confidence, provenance, and project scope. Lexical retrieval combines full-text relevance, recency, kind, confidence, and a bounded top-N result set. Say "remember this" to save explicitly.
Ten tools cover file reads, approved shell execution, approved writes, task state, checkpoints, and memory. execute_bash and write_file require a single-use confirmation token.
When the hosted deployment routes tool jobs to your paired agent, confirmed commands and writes execute on your machine inside a home-relative sandbox. Reads and data-plane tools do not use that second approval.
The dashboard aggregates attempts, failovers, and input/output token counts by day and provider/model. Provider-attempt history and continuity projections show how tracked work moved forward.
FAQ
Short answers with the same semantics the docs pin. Nothing here promises what the code does not do yet.
No. Create an account on the hosted service and everything — providers, sessions, memory, task state — lives server-side. Pairing a machine is one user command (pip install invincible-ai, then invincible agent). The same code also self-hosts on a laptop backed by PostgreSQL if you prefer to operate it yourself.
A 429 or 5xx records a failure, puts that credential in an exponential cooldown (30s up to a 300s cap), and tries your next credential. A 401 or 403 skips the credential. When every credential is exhausted the gateway answers 503 — your agent retries one stable endpoint instead of dying on the first 429.
On the hosted service, on your paired machine: the server keeps every security decision (denylist, staging, single-use confirmation token, audit, per-user routing) and the agent only does the work inside a home-relative sandbox, with denylisted credential paths blocked. An offline agent answers immediately instead of hanging.
Scoped lexical memories in PostgreSQL — notes, facts, preferences, decisions — ranked by full-text relevance, recency, kind, and confidence under one bounded budget shared with the continuity brief. Say "remember this" to save explicitly, and manage or delete everything from the dashboard. Semantic or vector search is not built yet.
Any OpenAI-compatible chat client against /v1/chat/completions, Claude Code against /v1/messages via ANTHROPIC_BASE_URL, Codex against /v1/responses, and any MCP-compatible client against /mcp through the browser OAuth flow.
No. Every request routes only through your own connected provider credentials — encrypted at rest, never shared across accounts. With no credentials connected, chat answers 400 and tells you to connect one on the Providers page.
No. Projects, credentials, sessions, runs, memories, continuity state, and MCP grants are all owner-scoped per account, verified by a closed multi-tenant audit — a foreign id reads exactly like an unknown one.
Run this on the machine you want to reach. macOS, Linux, and Windows — one package, no server to operate, no database to provision.
pip install invincible-ai invincible agent # browser opens → sign in → Approve