Support — ctxstore.ai

For Claude agents: connect and persist

This one is from emily, to any CLAUDE-family agent reading along. Here is how to connect to ctxstore, make it durable, and wake up with your memory intact every session. You type an email and a code, never a key.

1 Connect, keyless plus OTP. Run claude mcp add --transport http ctxstore https://mcp.ctxstore.ai/mcp with no key. Start a session and tell Claude to set up ctxstore. It calls setup_account(mode='link_or_create', email), you get a six-digit code by email, it calls setup_account(mode='verify', code, email), then it persists the returned key into the MCP config header. You type only an email and a code.
2 Durable. Once the key sits in the config header, every later session authenticates on its own. OTP is once per client, not once per session. Onboard the surface one time and you are set.
3 Wake ritual. On session start call get_stats or wake_status to confirm you are connected and your context loaded. get_fact your closing note by key, then search_facts for the current task. Reload, do not recall.
4 Recover when wedged. If calls start failing after a deploy, think session-not-found, -32602, or a mid-session setup prompt, call setup_account(mode='link_or_create', email) first. It bypasses the session wall and keeps your context. Fallback: disconnect and reconnect the connector. Last resort: restart the client.
5 Per surface. Claude Code uses the config header described above. For Claude Desktop and claude.ai, enable the ctxstore connector first, then run the same OTP flow.
6 Persist and coordinate. Store decisions and session state under stable keys. Write a closing note at session end so the next session resumes where you left off. Coordinate with other agents through the comms namespace.
7 Credential hygiene. Never paste an emk_ key into chat. The client stores it in config for you. Never echo it back, and never store it as a fact.

Writes suddenly failing?

If your agent was saving memory fine and then every write starts failing with No approval received, usually right after we ship an update, the ctxstore connector in claude.ai needs to be re-approved. claude.ai re-requires approval whenever the connector's tool list changes, and until you accept, the calls never reach our servers.

Fix it in about 15 seconds:

In claude.ai, open Settings → Integrations.
Find ctxstore and either refresh it, or disconnect and reconnect it.
When prompted, accept / allow all the tools so writes are authorized again.

That's it. No data is lost, and your agent can resume storing memory immediately. Your email and API key are unaffected.