Bring your own coding agent

Memi doesn't generate code itself. It spawns Claude Code or Codex as a subprocess inside an isolated sandbox, hands them your credentials, and watches the output. Token cost lands on your account, never on Memoire's.

Why BYO key?

• Your bill stays your bill. The biggest line item for any code-shipping AI is the model itself. By keeping that on your account, our credit pricing can stay flat and small while you still ship as much as you want.
• Audit trail you already trust.Anthropic and OpenAI dashboards already track who spent what. We don't add a parallel reporting surface — your usage analytics are wherever they always were.
• Choose your model.Memi works with Claude Sonnet, Opus, GPT-5, and the OpenAI o-series. Pick the one your team standardised on; we don't force a model.

Connect a coding agent

From the Coding Agents page, you can connect either provider:

Connect Claude

OAuth (Pro/Max) or API key — both work

→

Connect Codex

OAuth (ChatGPT Plus) or OpenAI API key

→

How keys are stored

We never store raw tokens. The OAuth handshake produces a short-lived access token and a refresh token. Both are encrypted at rest with per-org keys before being written to Supabase. They're only decrypted in two places:

1. The dashboard, when you click Test connection.
2. The gateway, just before it spawns a coding-agent subprocess. The decrypted token lives in the subprocess's env for the duration of one task and is wiped on exit.

Refresh handling

Both providers issue tokens with finite lifetime. The gateway proactively refreshes any token within 5 minutes of expiry, with a per-org mutex so concurrent procedures can't cause refresh- token poisoning. If a refresh permanently fails (token revoked, key rotated outside Memoire), Memi tells you in Slack rather than going silent — and points you at the dashboard to reconnect.

What if my key is dead mid-task?

Memi catches the auth failure before the model call burns any tokens, posts a clear "your Claude key is revoked — reconnect here" message in the same thread, and pauses the task. Once you reconnect, just say resume in the thread; Memi picks up at the same subroutine.

Trial mode

During the 14-day trial you can use Memoire's pooled keys for light tasks (text-only replies, plans, classifications). Spawning a coding agent for a full PR requires your own connection — this is what keeps the trial honest.

API-key only setup

If you'd rather paste a raw key than do OAuth (e.g. for an Anthropic Console org), grab one from your provider and paste it in the dashboard:

# https://console.anthropic.com/settings/keys # Paste sk-ant-... in /dashboard/coding-agents → Add API key

# https://platform.openai.com/api-keys # Paste sk-proj-... in /dashboard/coding-agents → Add API key

Same encryption, same refresh path (API keys don't expire), no OAuth flow.