OMI April 26 Update — what it means for Mark
Deep dive on OMI's April 9 + April 26, 2026 releases. Sourced from feedback.omi.me/changelog, OMI integration blog, GitHub release notes, and the deep-research agent run yesterday. Maps every change against your Oracle vision and TP3 stack.
1. What changed (the changelogs)
April 26, 2026 release — "3-Day Battery, Voice Replies, Desktop Task Execution"
- Claude 4.6 across all chat completion routes. The model behind OMI's chat answer generation. Was Perplexity for web search before. Now Claude across the board.
- BYOK (Bring Your Own Keys) for transcription. Route audio through your own Whisper / Deepgram / AssemblyAI key instead of OMI's default cloud.
- Voice replies — three modes: Off / Headphones-only / Always.
- Capture-screen tool added to MCP relay (now 14 tools on the relay; 8 tools documented on the public MCP server).
- Firmware 3.0.18 — 3-day battery. No hardware change required.
- Semantic search across action items.
- Desktop agent runtime — OMI desktop can now run tools end-to-end on your behalf.
April 9, 2026 release — "Phone Calls, Bluetooth Overhaul, Offline Sync"
- Web search swapped from Perplexity to Claude. Confirms the "it's using Claude now" you noticed.
- Phone calls with real-time transcription (Unlimited tier).
- Apple Reminders bidirectional sync.
- Auto Sync for Offline Recordings (firmware 3.0.17 beta).
- 15 new transcription languages.
- ~5-minute voice messages on mobile.
Earlier 2026 releases (still relevant to your stack)
- Feb 2026 — Omi Glass audio + video, OTA firmware updates, Apple Health (active energy + sleep) integration, 11 new third-party integrations (Slack, Notion, Dropbox, GitHub, Spotify, etc.).
- Jan 2026 — Folders, Daily Recaps (auto at 10 PM), web app at home.omi.me, Goals, 3D brain map, free unlimited on-device transcription via developer settings.
2. What it means for your Oracle vision
Recall the Oracle north-star: trigger word ("Computer,") → OMI captures the question → OMI runs Claude with context (OMI memories + web search + TP3 facts) → answer delivered via push notification, voice reply, or in-app chat.
Before this update, that flow had three gaps. The April 26 release closes two of them and gives you tools to close the third.
| Oracle gap | Pre-April-26 state | Post-April-26 state |
|---|---|---|
| Answer quality | Web search via Perplexity, chat via older model | ✓ Claude 4.6 across both paths. Auto-rolled. |
| Voice output | Push notification only — silent | ✓ Voice replies through headphones. Needs your toggle. |
| Memory bidirectional | OMI captures audio → TP3. Nothing the other way. | ✓ MCP create_memory exposed. Built it last night — see §4. |
| Cost control | OMI's default cloud transcription on their key. Last time we had a silent fallback path it cost $60/mo. | ✓ BYOK transcription. Costs land on a key you set the cap on. |
3. What auto-rolled (no setup needed)
4. What needs your hands (3 things, ~10 min total)
5. What Claude built last night (the bidirectional piece)
The biggest opportunity in this update was the MCP create_memory tool — OMI now lets external agents WRITE to its memory store, not just read.
What's now wired:
tp3_claude_to_omi_sync.pyat/home/g16/.claude/scripts/— scans Claude's auto-memory directory, pushes each new/modified memory file as a private OMI memory taggedclaude-memory+ type. State file dedupes by mtime so nothing pushes twice.- 5 memories already pushed live last night — you verified the format looked right.
- Run nightly via cron / scheduled task (TBD; manual for now).
Net effect for Oracle: when you say "Computer, what did I lock in about v3 pricing?" OMI's chat now has access to feedback_legacy_soil_informational_not_pitch_2026-04-26.md alongside its own conversation memories. Oracle becomes Mark-aware, not just OMI-aware.
6. New integrations worth flipping
| Integration | Why it matters for you |
|---|---|
| Apple Reminders | Oracle can create tasks that show up in your phone's Reminders app. Shortcut: brain-dump action items, OMI auto-creates reminders. Bidirectional — items you check off in Reminders update OMI. |
| Phone calls + transcription (Unlimited tier) | Real-time transcription of your phone calls. Useful for the contract meetings (Brian Daly), vendor calls, parent calls. Auto-transcribed, auto-stored, queryable. |
| Slack / Notion / Dropbox / GitHub (Feb 2026 batch) | If you ever do GitHub-side work for the projects, OMI can search your repos. Notion useful if you migrate any docs there. Skip Slack unless you join a community channel. |
| Apple Health (Feb 2026) | Active energy + sleep into OMI context. Oracle can answer "how did I sleep last night?" or "have I been moving enough this week?" |
| Folders (Jan 2026) | Organize OMI memories by project. You could create a "Legacy Soil" folder, a "Bidet" folder, a "School" folder for cleaner Oracle queries. |
7. What to skip (don't adopt)
8. Pricing / plan / hardware
- Your plan: Unlimited Monthly $19.99/mo, renews October 24, 2026.
- Device: $89 one-time (already owned).
- No new hardware in this release. DevKit 2 still current. Brain-interface module still deferred to 2026-2027.
- Free tier still includes: unlimited on-device transcription + 1,200 cloud minutes/month.
- Phone calls + full Claude/MCP feature set are gated behind Unlimited (you have it).
9. Cost guardrails (lessons from the Gemini fallback incident)
April 22 we learned that silent cloud fallbacks compound silently — Bidet quietly hit Gemini fallback when Ollama died for days, blew $60/mo before we caught it. The April 26 OMI update gives you the tools to never let that happen again on the OMI side:
- BYOK transcription — costs go to a provider account YOU control. Set a $15/mo cap on the provider side (Deepgram Console → Spending Caps). Provider emails you at 50% / 80% / 100%.
- 14-tool MCP relay — OMI tells you what tools fired and when. Auditable.
- Folder organization — quarantine experimental flows so they don't pollute the main memory store.
10. Concrete action items, ranked by impact
| # | Action | Effort | Why |
|---|---|---|---|
| 1 | Backfill ALL Claude memories to OMI: python3 /home/g16/.claude/scripts/tp3_claude_to_omi_sync.py --max 100 | 30 sec | Pushes every standing rule + project state + reference into Oracle's context. Big jump in Oracle accuracy. |
| 2 | Try Oracle: "Computer, summarize what Claude shipped yesterday" or "Computer, what's the status of Legacy Soil v3?" | 1 min | Real-world test of the bidirectional sync. If Oracle answers from Claude memory content, the sync is working as intended. |
| 3 | BYOK Deepgram setup | 5 min | Cost control + audit. Eliminates the silent-fallback risk that bit you on Gemini. $15/mo cap. |
| 4 | Voice replies → Headphones-only | 30 sec | Turns Oracle voice live without making it socially awkward. |
| 5 | Apple Reminders bidirectional | 1 min | Oracle action items auto-flow into your phone Reminders. Closes the loop on "what does Mark need to actually do today." |
| 6 | Install Claude→OMI sync as nightly scheduled task | 5 min | Make the bidirectional sync auto-run instead of manual. Each new memory you save lands in Oracle's context within 24 hours. |
| 7 | Verify firmware 3.0.18 (3-day battery) | 30 sec | Settings → About → Firmware version. If you're on 3.0.16 or older, it'll auto-update. |
| 8 | Folders: create "Legacy Soil," "Bidet," "School," "Personal" | 2 min | Cleaner Oracle queries. "Computer, what's in my Legacy Soil folder?" |
11. What this doesn't change
- TP3 ingest pipeline — webhook chain (Conversation Events / Real-time Transcript / Audio Bytes / Day Summary → Cloudflare Worker → Apex ingest) is unchanged. OMI still pushes to TP3 the same way.
- Bidet AI — separate app, separate workflow. The Gemini fallback in Bidet stays disabled. Bidet now runs on GPU faster-whisper, no cloud transcription.
- Apex's role — still the always-on Windows server hosting TP3 + Bidet + omi-mcp SSE relay + scheduled tasks.
- OMI's role — still the capture + chat + Daily Recap layer. Doesn't replace Apex; complements it.
12. The "what could change" longer-horizon angle
Two trajectories worth tracking on the OMI roadmap as they could materially change your stack:
- Brain-interface module (2026-2027 horizon). If it ships, OMI moves from voice-capture to subvocalization / direct neural input. That's a different category of capture and would need a re-architecture of how it feeds TP3.
- OMI Desktop agent runtime maturity. If it gets robust enough to genuinely replace Apex's scheduled-task layer, you could simplify the stack — kill Apex, run everything off OMI desktop. Years away from production-grade for your use case, but worth tracking.
Compiled 2026-04-26 PM from OMI deep-research agent run + last night's reframe conversation. Sources: feedback.omi.me/changelog, omi.me/blogs/integrations, docs.omi.me/doc/developer/mcp/setup.md, github.com/BasedHardware/omi, internal memory reference_omi_platform.md.