Troubleshooting
Most issues fall into one of five buckets. Skim the headings to find yours.
My AI app says it can't reach Lune
If the connection fails outright (timeout, "couldn't reach the server"):
-
Confirm the URL. Remote MCP lives at
https://mcp.luneresearch.com/mcp. Some apps want you to drop the/mcpsuffix; most don't. Try both if the first form fails. -
Confirm your network can reach it. From a terminal:
curl -sS -o /dev/null -w "%{http_code}\n" \ https://mcp.luneresearch.com/.well-known/oauth-protected-resourceYou should see
200. Anything else points at a network or DNS issue on your side. -
Check our status page (if linked from the dashboard) or the dashboard banner for an active incident.
I get "401 unauthorized" or "invalid token"
The token your AI app is sending doesn't match a live key.
- Personal access token: open the dashboard's
Credentials page,
generate a fresh token, and paste it back into the AI app's connector
configuration. PATs start with
lune_. Free-tier tokens never expire, but anyone with dashboard access can revoke them. - OAuth (browser sign-in flow): sign out and sign in again in the AI app's connector settings. The token refresh path is automatic, but a long-idle session can drift past the refresh window.
- Wrong team selected: a token belongs to one organization. If you recently switched teams in the dashboard, the old token still points at the old team. Mint a new token under the team you want to use.
I get "rate limited" or "429"
You've hit the daily request cap or run out of credits.
- Free plan caps you at 10 requests per day, resetting at UTC midnight. Pro raises the cap to 300, Max to 600. Upgrade from Settings → Billing.
- Credit packs ($0.99 for 300 extra requests) keep you working past the daily cap. Top up from the same Billing page; credits never expire and pool across the team.
- Burst limits: we also cap raw burst at 30 requests per second to guard against runaway scripts. If your AI app fans out a lot of tool calls in parallel, the burst limiter may kick in even before the daily cap. The error message tells you exactly which limit fired.
My AI app connects but the tools don't show up
The MCP session opened, but tools/list returned nothing or the app
hasn't refreshed its tool catalog.
- Restart the AI app. Most clients only fetch
tools/liston cold start. - For ChatGPT and other web AI apps, remove the connector and re-add it. Web clients sometimes pin a stale session ID after a deploy on our end.
- For Claude Desktop, quit fully (not just close the window) and relaunch.
- Confirm your token has the right scopes. The default PAT has all four
scopes (
papers:read,guidance:read,subs:rw,account:read); a custom-scoped key may exclude tools you expect.
For deeper, app-specific tips, see Desktop AI apps and ChatGPT and other web AI apps.
Search results look wrong, missing, or outdated
- Coverage: Lune indexes the top peer-reviewed venues
(NeurIPS, ICLR, ICML, ACL, EMNLP, CVPR, ICCV, USENIX Security, IEEE
S&P, CCS, NDSS, and more). If a venue isn't on that list, the paper
won't be in our corpus. The full list is available from the
list_conferencestool. - Freshness: new accepted papers reach the corpus within 24 hours of the conference's official posting. If you're seeing yesterday's acceptances missing, check back tomorrow.
- Phrasing: Lune does best with full natural-language queries ("methods for retrieval-augmented generation that reduce hallucination on long-form QA"). Short keyword bags ("RAG hallucination") will often miss relevant papers.
- Conference filters: short-name matching is fuzzy but strict. If your filter doesn't match, the search falls back to no filter rather than silently corrupting results. The error message lists candidate short names when there's ambiguity.
I'm still stuck
Email [email protected] with:
- The AI app you're using (Claude.ai, Claude Code, ChatGPT, Cursor, etc.)
- The exact error message
- A
request_idif your AI app surfaced one (every Lune response carries one in the response data) - Whether it's a brand-new install or a previously-working setup
We typically reply the same day.