Quickstart

Portable cndk_v1 key (cross-vendor)

A personal key gives you one memory across many clients. This is the path the "any LLM" story rests on. Every example below uses the placeholder cndk_v1_…. Never paste a real key into a page or a shared file.

Step 1 — Mint your key

1. Open the dashboard at /cortex-dashboard.html on app.conduir.ai — key generation is at the top of the page.
2. Sign in with Microsoft. Minting is session-gated.
3. Click Generate key.
4. Copy the secret now. It is shown once and never stored on the page. The key is bound to your Microsoft identity, not a shared demo user.

Step 2 — Know the broker contract

Every client speaks to the same broker with the same shape.

Step 3 — Connect a client (once)

Wire the key into a client one time. After that the agent recalls and remembers on its own — you are not sending requests by hand. The curl in Verify below is only a smoke test.

Claude Code is the fastest: one command straight at the broker with your key, no local process and no OAuth.

claude mcp add --transport http cortex https://app.conduir.ai/api/cortex/mcp \
  --header "Authorization: ApiKey cndk_v1_…"

claude mcp list        # => cortex … - Connected

Gemini CLI — point it straight at the broker over HTTP. No local process.

npm install -g @google/gemini-cli

gemini mcp add --transport http --trust \
  --header "Authorization: ApiKey cndk_v1_…" \
  cortex https://app.conduir.ai/api/cortex/mcp

gemini mcp list        # => cortex … (http) - Connected

ChatGPT (web / desktop) — via a Custom GPT Action.

1. In a Custom GPT under Actions, add one POST operation to the public broker (the endpoint in Step 2) with the flat {tool, arguments} body from the curl examples above — no spec file to import.
2. Set Auth to API Key, type Custom, header Authorization, value ApiKey cndk_v1_….
3. It calls the public broker directly, with no local process.

OpenAI Codex CLI — direct over HTTP with the key. (Codex also supports OAuth — see Path B.)

# ~/.codex/config.toml
[mcp_servers.cortex]
url = "https://app.conduir.ai/api/cortex/mcp"
http_headers = { "Authorization" = "ApiKey cndk_v1_…" }

Your own app — wrap the broker once, then call it like any function. No shim, no per-call curl.

// cortex.js — write this wrapper once
const BROKER = "https://app.conduir.ai/api/cortex/mcp";
export async function cortex(tool, args) {
  const res = await fetch(BROKER, {
    method: "POST",
    headers: {
      "Authorization": "ApiKey " + process.env.CONDUIR_API_KEY,
      "Content-Type": "application/json"
    },
    body: JSON.stringify({ tool, arguments: args })
  });
  if (!res.ok) throw new Error("cortex " + tool + " -> " + res.status);
  return (await res.json()).result;
}

// then anywhere in your app, no HTTP boilerplate:
await cortex("memory.remember", { content: "user prefers dark mode", idempotency_key: crypto.randomUUID(), content_kind: "fact" });
const hits = await cortex("memory.recall", { query: "user preferences" });

Grok CLI — point it straight at the broker over HTTP with the key. No local process. The broker side of this is verified; a Grok session over it has not been re-verified end-to-end, so treat it as you would the Gemini row above.

npm install -g @xai-official/grok

grok mcp add --transport http cortex https://app.conduir.ai/api/cortex/mcp \
  --header "Authorization: ApiKey cndk_v1_…"

grok mcp list        # => cortex … - Connected

Claude Code over OAuth (no key, ~60s)

One command wires Cortex into Claude Code over MCP. No SDK and no glue code.

1. Add the memory MCP server:
   Copy

   claude mcp add --transport http memory https://app.conduir.ai/memory --client-id cortex-claude-code

2. Run /mcp, sign in in the browser, and it connects. The install flow is add → /mcp → sign in → connected.
3. The agent starts remembering on its next turn.