Reported issues for Diagrams.so

Pod holds 8 of 8 problems reported by people outside the maintainer team. Issues filed by the project's own owners, members and collaborators are excluded entirely — a maintainer's release checklist is not a warning to a prospective user.

Back to Diagrams.so.

Most discussed

MCP go-live: exact steps to ship v1.3.0 to Claude Code / Claude Desktop / Cursor users

Tracking ticket for launching the Diagrams.so MCP server so end users can install it from Claude Code, Claude Desktop, and Cursor. Ordered by hard dependency — steps 1–4 are the critical path; step 7 runs in parallel after step 2.

Updated 2026-08-03. This ships v1.3.0 (the v1.2.0 tool surface plus the audit-fix release), not v1.2.0. Every state below was verified live against GitHub + production.

Where we are (verified 2026-08-03)

• Public API live at `https://api.diagram

Read the thread · 2026-07-24 · open · 2 comments

Session credit tally under-reports: failed-but-billed calls invisible, re-layout charges never counted

From the 2026-07-30 session audit (M3) — verified by call-site inventory:

What actually happens: the in-process tally recorded only successfully-parsed responses of generate/edit/fix. A call that failed mid-flight (e.g. the audited 504 that WAS billed server-side) was silently absent, and re-layout charges were never recorded at all — so the tally and the server ledger showed contradictory numbers side by side in the same get_usage_history output, with no caveat.

Fix (v1.3.0): ambiguou

Read the thread · 2026-08-01 · closed · 1 comment

Retrying a billable call after a timeout/504 double-bills — no Idempotency-Key sent; client timeout sits below the server's worst case

From the 2026-07-30 session audit (M2, client half) — verified, then reproduced against a stub API:

What actually happens: the server supports Idempotency-Key (24h replay window), but v1.2.0 never sent one (zero occurrences in the source). When a long generation outlives the edge timeout, the server completes AND charges while the client sees 504 with no diagram id — and any retry is a brand-new request: second diagram, second charge. The 180s default client timeout also sat below the ser

Read the thread · 2026-08-01 · closed · 1 comment

In-tool auto-connect (beginConnect) still sends the pre-email device-code request — every headless connect gets 426 DEVICE_CLIENT_OUTDATED

What breaks

beginConnect() in src/client.ts (~L266) posts POST /oauth/device/code with {client_id, livemode, device_name} — no email. Since the API's emailed-code redesign went live (~Aug 8, app-core 72e98f3), that request is always rejected 426 DEVICE_CLIENT_OUTDATED. It also still expects json.user_code and json.verification_uri_complete, which the new contract deliberately no longer returns — so even an accepted request would fail the parse.

The interactive path (`c

Read the thread · 2026-08-19 · closed · 0 comments

Go-live: Claude connector — remote MCP server over Streamable HTTP (additive; no migration)

Claude connector — remote MCP server over Streamable HTTP (go-live)

Ships the resource half of the Claude connector: a stateless remote MCP server that speaks the MCP Streamable HTTP transport, does the 401 + WWW-Authenticate handshake that makes Claude show the Connect card, and forwards each user's OAuth Bearer to the Diagrams.so API per request. Paired with the OAuth authorization server + consent page in diagramz-app-core (separate PR/issue).

Additive. The existing **std

Read the thread · 2026-08-14 · open · 0 comments

Go-live: emailed-code login — no migration, release after backend

Tracking issue to ship the anti-phishing login change (PR #13, draft).

Summary

login now prompts for an email (or DIAGRAMS_LOGIN_EMAIL), sends it so the backend emails a one-time code, opens the plain /device page (no code in the URL), and calls /device/confirm after saving the key (delivery-safe rotation — no lockout).

Migration — NOT required

Client only. No schema changes.

⚠️ Release order — do NOT publish before the backend is live

This client calls the new emailed-code

Read the thread · 2026-08-05 · open · 0 comments

Publish to the official MCP registry (and fix the npm metadata that directories read)

We are absent from registry.modelcontextprotocol.io. The only result for "diagrams" today is a third party, io.github.mskry/diagrams-mcp-server. That is the listing an agent reads when asked what MCP servers exist, and the highest-authority one available to us.

Why it never got filed

The documented route is mcp-publisher login github, which runs its own device flow and needs a human at a browser signed in with RedHold org rights. That is a manual step nobody owns, so it kept not happe

Read the thread · 2026-08-04 · closed · 0 comments

Wrong settings tab in four launch-surface READMEs, and a shorter CLI command

Two things, both affecting what a brand-new user sees first. Raised from the engineering asks of 4 Aug, verified against main and the running app.

1. Four READMEs send people to a tab that does not exist

The settings tab is labelled AI Provider (apps/web/src/app/settings/page.tsx:657), and the API keys section lives inside it. There is no "API Keys" tab.

Read the thread · 2026-08-04 · open · 0 comments

Most recent

The remaining reports are on the project's issue tracker.