August 2026 Release Notes: A Desktop App, Interactive MCP Apps, and Claude Opus 5

TL;DR

August was about extending Evermuse beyond the browser tab - onto your machine and into your AI client - while tightening up the plumbing underneath. Everything below is new this month. The shortlist:

• A desktop app: Evermuse now runs locally - an Electron app with its own local MCP server and on-device SQLite, v0.1.0 and deliberately minimal for now.
• Interactive UI inside your AI client: MCP Apps let tool results render as themed, chrome-free views instead of raw text - live ingestion progress, a source catalog, and more, gated in Labs.
• A leaner MCP tool surface: Search now returns relevance scores and paginates, find_skills returns a first-run context bundle, and every tool call is certified before it runs.
• Import from anywhere: Bring in external meetings and communications through add_source, deduped by vendor identity and resumable if interrupted.
• A friendlier MCP-first signup: Sign up straight from Claude, Cursor, or another MCP client and land in a real workspace instead of a dead end.
• A more honest meeting pipeline: Failed processing now says so instead of silently marking a meeting complete, with a health-check cron watching for stragglers.
• Claude Opus 5 and Gemini 3.6 Flash: Opus 5 is the new default chat model, and Gemini 3.6 Flash joins the selector.

🖥️ A desktop app, running locally

Evermuse now has a desktop app. A new desktop/ workspace bootstraps an Electron app paired with a local MCP server and on-device SQLite, so a meaningful slice of Evermuse can run without a round trip to the cloud.

• Its own local MCP server. The desktop app ships a local MCP server, so your AI client reaches Evermuse the way it does today - just running on your machine.
• On-device SQLite. Data the desktop app manages lives in local SQLite instead of round-tripping to the backend for every read.
• Local-first workspaces, more broadly. Workspaces outside production can now opt into store_locally at creation, which bifurcates the MCP tool surface and built-in skill set for that workspace - the same local-first posture that powers the desktop app, available without installing it.
• Early days, on purpose. This is v0.1.0: no onboarding UI yet, no embeddings, and only find_sources from the local tool set. macOS builds are available now.

🎨 Interactive UI inside your AI client

MCP tool calls used to return raw text or bare HTML. A new (Labs-gated) render_ui router now serves proper ui:// MCP App resources, so results can render as real interfaces inside your AI client.

• Themed, chrome-free views. Raw agent HTML views are replaced with themed React templates that match Evermuse’s look, with no browser frame around them.
• A source catalog with real icons. A deterministic, icon-backed catalog replaces guesswork about what kind of source you’re looking at.
• Live ingestion progress. Watch an import happen in real time, with post-ingestion slash-command options and a list-capabilities picker surfaced right in the UI.
• Gated behind Labs. The new mcp-apps-ui flag ships seeded off - dark for everyone until a workspace turns it on.

🔧 A sharper MCP tool surface

A round of consolidation on how AI clients talk to Evermuse over MCP.

• Streamlined tools and skills. The third-party MCP tool surface, skill set, and call path got a pass, with follow-up rounds covering filter pagination, cache TTL, token-family revocation, and search-batch as the one required grounding step.
• Better search. search now returns absolute relevance scores, paginates, and leads with a digest.
• Smarter skill discovery. find_skills returns a first_run context bundle and deduped citations instead of a flat list.
• Certified before it runs. Every tool call is now gated on a skill_safety_check, and tool results carry secure deep links.
• Faster key checks. API-key lookups can opt into a short in-memory cache, skipping a database round trip on hot paths.

📥 Import conversations from anywhere

add_source now covers more ground.

• External meetings and communications. Import them through add_source, deduplicated by vendor identity, and resumable if the import gets interrupted partway.
• Ingested emails render properly. Emails you bring in now display in the single-source view instead of as raw text.
• Unstranded MCP imports. Meetings imported via MCP get signed AV URLs, loud failures instead of silent stalls, and correct source labels.
• A spreadsheet importer, in Labs. Round-trips a real knowledge-base export, gated behind a new flag that’s off everywhere until a workspace opts in.

🔐 A friendlier MCP-first signup

Signing up for Evermuse straight from an MCP client is smoother.

• No more dead ends. MCP-first signups are onboarded and tagged instead of getting stuck with no workspace.
• A consistent consent screen. The OAuth consent flow was restyled to match regular sign-in.
• A guided handoff. Setup now hands off to the setup-evermuse skill, so your MCP client walks you through the rest.
• Microsoft redirect fixed. A full-page Microsoft redirect now returns you to the MCP consent flow instead of dropping you on the dashboard.

🧵 A more honest meeting pipeline

Fixed a class of failures where a broken meeting looked identical to a successful one.

• Failures say so. Terminal pipeline failures now write an explicit error status instead of silently flipping to “completed” - previously the root cause of bot meetings sitting with transcripts but no signals or overview, undetected.
• A health-check cron. A new job runs every 15 minutes, flags stalled or errored meetings, and alerts once per meeting instead of repeating.
• Consistent status everywhere. The frontend’s status badges had drifted from the backend’s, so failed meetings showed no badge at all in the sources sidebar. Both now share one status-classification module.
• Fixed upload ordering. Uploads no longer get marked “completed” before their overview actually exists.

💳 Billing: fewer dead ends

Running out of credits, or just checking usage, got less confusing.

• Out-of-credits workspaces see a path forward. Instead of failing opaquely, they’re shown the upgrade path directly.
• Usage page gets billing-period navigation. Move between billing periods instead of only seeing the current one.

🤖 New models: Opus 5 and Gemini 3.6 Flash

• Claude Opus 5 is the new default. Added to the chat model selector and made the default model; Opus 4.8 is now marked legacy.
• Gemini 3.6 Flash added. Joins the selector as an older preview model is retired outright.
• A pre-deploy safety net. Shaping’s thinking budget is now clamped below each model’s output cap, and a new model smoke suite runs as a hard gate on every dev, staging, and prod backend deploy.

✨ Polish

• Citation actions no longer create phantom page overflow on the project page.
• Internal voices can no longer be quoted as customer evidence by the daily-brief skill.
• Dark-mode contrast fixed in the roadmap opportunity dialog; monochrome dev-tool logos inverted for dark mode.
• A costly Firestore delete path no longer serializes hundreds of reads.
• Two logging defects fixed, and a set of dependencies patched for known security advisories.
• Bash can no longer read secret config files.
• 45 unused frontend files removed, 2 components deduped.
• A round of flaky-test and CI reliability fixes, plus dozens of new unit tests across DAOs, ingest routes, and PII scrubbing.

For questions, feedback, or support, please reach out to the Evermuse team at team@evermuse.com - we’re here to help!