Magic2 · meeting intelligence · v1.0shipping today

There’s nothing else like this.

Otter takes notes. Fireflies records audio. Granola writes a summary. Magic2 does all that — and then builds you a live mind map, generates eleven separate deliverables the moment you stop, remembers across every conversation you’ve ever had, and renders the whole thing as a magazine-quality PDF. Locally. On your own GPU.

11
artifact kinds
action items · exec summary · slide deck · MVP spec · decision log · risk register · interview · hiring brief · investment memo · legal memo · contract redline
26
meeting templates
from standup to design review, vendor pitch, hiring loop, customer discovery, legal review, retro — Magic2 knows the shape of your meeting
384d
vector recall
MiniLM-L6 embeddings on every transcript chunk and every topic node. Cmd-K finds anything you've ever said.
1.0s
end-to-end latency
from word spoken to word on screen. WhisperLive over WebSocket. No buffering, no batching.
preview
Live mind map mid-meeting
Live mind map mid-meeting
Live nodes
Topics, decisions, risks, questions, actions — extracted every 15 seconds.
Meeting root
Every map anchors to a gold root: meeting title + time.
Affinity edge
Semantic similarity links between near-duplicate ideas. Auto-deduped.
Ghost subgraph
Cyan dashed nodes are prior-meeting topics. Continuations surface as you talk.
§01 · the sixty-second pitch

Here’s what happens
in the first minute.

Most meeting tools wait until the meeting is over to do anything useful. Magic2 starts working the moment you press record — and keeps working through every second of the conversation.

  1. T+00:00

    You click record.

    Microphone permission, one prompt. The WebSocket opens to your local backend, which opens its own to WhisperLive on a GPU box on your LAN.

  2. T+00:01

    Words land on the screen.

    Transcription is live, not buffered. You see your sentence form as you finish it, with about one second of glass-to-glass latency.

  3. T+00:15

    The first nodes appear.

    qwen2.5:7b-instruct emits the first topics, decisions, risks, questions and actions. Each lands anchored to the second it was said.

  4. T+00:30

    Continuations surface.

    If this meeting is the natural sequel to one you had three weeks ago, a shooting star arcs across and the prior topics ghost in.

  5. T+01:00

    Magic2 names the meeting type.

    Design review? Vendor pitch? Customer discovery? Magic2 picks one of 26 templates and starts watching for the items each one needs.

  6. T+n.00

    Stop. The work is at the door.

    Eleven artifacts render in parallel. A smart title arrives two seconds later. Cross-meeting links score. Cmd-K opens.

§02 · the eleven

11 deliverables. One conversation.

The moment you stop the meeting, Magic2 renders eleven separate documents in parallel. Each one carries source-anchored quotes back to the moment they came from. None of these is a generic “summary”.

A-01 · artifact

Action Items

Owner, due date, evidence quote, link back to the timestamp. Sortable, exportable, ready for Linear or Slack tomorrow.

action_items.md · A-01
meeting · API redesign · 5 items · 3 with owners
  • BobWrite rate-limit runbook · pin to on-callFri
  • MiaSpike GraphQL gateway alternative · 1 dayMon
  • LinaDraft RFC for v2 auth flow · circulate SlackMay 28
  • WillCost-model the migration · vendor sheetMay 24
  • DevsReplace deprecated JWT library across servicesQ3
source: transcript 142.4s—167.9sexport → linear · slack · csv
A-02 · artifact

Executive Summary

A single page. Headline takeaway up top. Five bullet outcomes. A ‘what’s next’ line. Built for the people who didn’t attend.

executive_summary.md · A-02
★ key takeaway

The team chose REST over GraphQL for the new public API, prioritising cache simplicity and edge familiarity over query flexibility.

outcomes
  • ·Public API spec to be drafted by Friday (Lina)
  • ·Vendor scoping limited to two finalists; final cost model by May 24
  • ·Existing JWT library scheduled for deprecation across services in Q3
  • ·Open question: rate-limit policy per consumer tier — runbook due Friday
what’s next

RFC circulated by Mia early next week; consensus check at Friday standup; spec freeze before sprint cut.

A-03 · artifact

Slide Deck

A talking-points deck — title slide, agenda, one slide per topic, a closing call-to-action. Open it in Keynote and you’re ready.

deck.pdf · A-03
12 slides · 16:9 · ibm plex
slide 01
API Redesign
May 19 · Design Review
magic2
slide 02
Agenda
1. Why · 2. Options · 3. Decision · 4. Risks · 5. Actions
magic2
slide 03
Why now
Public API requests up 4×; cache strategy not scaling.
magic2
slide 04
Options
REST · GraphQL · gRPC-Web
magic2
slide 05
Decision
REST. Cache simplicity wins.
magic2
slide 06
Next
RFC · spec · vendor cost model
magic2
A-04 · artifact

MVP Spec

If the meeting was about a product, Magic2 emits a feature matrix with priority, dependency, and evidence quotes per row.

mvp_spec.md · A-04
IDFeaturePriorityDependsEvidence
F-01Token-bucket rate limit per consumermust142.4s
F-02Webhook delivery with at-least-oncemustF-01203.2s
F-03OpenAPI 3.1 spec auto-publishedmust289.7s
F-04API key rotation UI in dashboardnextF-01412.0s
F-05Per-customer audit log exportnextF-02511.8s
F-06GraphQL gateway adapterlaterF-03
A-05 · artifact

Decision Log

Every decision the room made, with the rationale, dissenters, and a transcript quote. Audit-ready.

decision_log.md · A-05

REST over GraphQL

May 19
why
Cache simplicity · edge familiarity
quote
“…we’ll get further with REST and a CDN than fighting cache invalidation.”

Deprecate legacy JWT library

May 19
why
Security advisory · maintenance load
quote
“the lib is unmaintained, that’s reason enough.”

Two-vendor finalists only

May 19
why
Cap evaluation cost · clear comparison
quote
“three is one too many for a four-week timeline.”
A-06 · artifact

Risk Register

Risks with impact, likelihood, owner, mitigation. Heatmap on the side. Source quote per row.

risk_register.md · A-06
IDRiskImpactLikelihoodOwner
R-01Migration breaks v1 consumers
Lina
R-02Rate limit too aggressive for partners
Bob
R-03Vendor lock on cache provider
Will
R-04Auth migration concurrent with API cut
Mia
A-07 · artifact

Customer Interview

Themes, pain points, jobs-to-be-done, verbatim quotes. The output a product researcher actually wants.

customer_interview.md · A-07
themes
onboarding frictionpricing transparencyslack integrationdata exportteam admin
pain points
  1. “Took us a week to invite the team because admin and billing are different surfaces.”
  2. “We don’t want another inbox; we want it in Slack where we already triage.”
jobs-to-be-done

“When my team finishes a sales call, I want the next steps in our CRM before I’ve walked back to my desk, so I don’t have to chase anyone.”

A-08 · artifact

Hiring Brief

Role profile, must-haves, nice-to-haves, the bar to clear, sample interview questions — derived from the calibration conversation.

hiring_brief.md · A-08

Senior Platform Engineer

L5 · remote OK
must-haves
  • ·5+ years building public APIs at scale
  • ·Owns rate-limiting / quota systems end-to-end
  • ·Comfortable writing RFCs and getting them through
bar to clear

Has shipped a public API that survived 10× traffic growth without redesign. Can explain why cache-first beat them at some point.

A-09 · artifact

Investment Memo

Thesis, market, team, traction, risks, ask. Structured the way LPs want to read it. With evidence.

investment_memo.md · A-09

Investment thesis

Live meeting intelligence is the ‘Notion moment’ for unstructured workplace conversation: the value isn’t the transcript, it’s the deliverable on the other side.

Market
Knowledge workers · TAM ~$24B
Team
Sole founder · ex-eng leadership
Traction
Beta · 28 dev users · NPS 71
ask
$1.5M pre-seed · 18-month runway · two senior hires · GTM motion
A-10 · artifact

Legal Memo

Issue / analysis / conclusion / recommendation. Disclaimer banner is hard-coded — the LLM cannot remove it.

legal_memo.md · A-10
⚠ disclaimer
AI-generated. Not legal advice. Verify all assertions with qualified counsel before relying on this document.
issue

Whether the proposed indemnification clause shifts unreasonable risk to the supplier.

analysis

The clause as drafted lacks a cap, materially exceeding industry norms (typically 1–2× annual fees). Negotiating leverage suggests…

recommendation

Insert a cap at 2× ARR; carve out IP infringement to remain uncapped.

A-11 · artifact

Contract Redline

Before / after diff, clause-by-clause, with the rationale for each change pulled from the discussion.

contract_redline.md · A-11
⚠ disclaimer
AI-generated redline. Have counsel review before signing.
§ 8.2 indemnification
-Supplier shall indemnify Buyer for any and all losses, damages and expenses arising out of any breach.
+Supplier shall indemnify Buyer for direct losses, capped at two (2) times the annual fees paid; provided, that claims arising from third-party IP infringement shall remain uncapped.
why

Discussed at 14:32 — team raised concerns about uncapped exposure exceeding standard market terms.

§03 · field comparison

What your current tool does.
And what Magic2 does on top.

Note transcribers stop at the transcript. Summary tools stop at the summary. Magic2 starts where they finish.

CapabilityMagic2Otter / Granola / Fireflies
Live transcriptionyes · ~1s latencyyes (most)
Mind map that grows as you talkyesno
11 deliverables generated post-meetingyessummary only
Cross-meeting semantic recallyes · 384d vectorsno / keyword only
Continuation detectionyes · autono
3D meeting constellationyesno
Magazine-quality PDF exportyes · IBM PlexPDF of summary
Local LLM (your GPU)yes · defaultno · cloud API
Audio never leaves your networkyes · defaultno · cloud upload
26 meeting type templates with coverageyes · liveno
Smart auto-namingyes · post-stopmanual / date
Source-anchored quotes in every artifactyessometimes
Cmd-K semantic search across all historyyesno / keyword
§04 · anatomy of the mind map

Every dot is a moment.

The mind map isn’t decorative. Each node carries a kind, a label, a summary, a start and end timestamp, and edges to its parents. Eight node kinds cover the shape of any conversation.

Q3 PLANAPI redesignHiringREST > GraphQLBob: runbookCache riskWho owns RFC?Bob(ghost: vendor)
  • topicwhat the room is talking about
  • decisionsomething the room decided
  • actionsomething someone will do
  • risksomething that could go wrong
  • questionsomething the room is still figuring out
  • entityperson, company, product, place
  • ghostnode from a prior meeting (dashed)
  • rootthe meeting itself (gold)
§05 · cross-meeting recall

Magic2 never forgets
what was said.

Mid-meeting, mention a topic you discussed three weeks ago — a shooting star arcs across the map and the prior conversation surfaces, ghosted in cyan, anchored to the moment that triggered it.

Entity JaccardShared people · companies · products
Node overlapShared topics / decisions / risks
Centroid cosineEmbedding similarity across full meeting
Recency decayRecent meetings weigh slightly higher
score = 4·entity_jaccard + 3·node_overlap + 2·centroid_cos + 1·recency
M-2024-04-19 · priorM-2026-05-19 · nowSCORE8.7/10
preview
3D meeting constellation
3D meeting constellation
gold = meeting root
violet = topic cluster
cyan = link strength
§06 · 3d constellation

Every conversation,
in one sky.

Spin the constellation. Meetings that talk about the same things drift toward each other. Edges thicken with entity overlap, centroid similarity, and recency. Hover a star — see who, what, when, score.

We’ve looked. No other meeting tool has anything like it. Granola has a list. Otter has a list. Magic2 has a galaxy.

§07 · cmd-k semantic search

Every word
you’ve ever said.

Press ⌘K. Type a phrase. Magic2 searches every transcript chunk and every map node you’ve ever recorded — with embeddings, not keywords. Find the exact quote without remembering the meeting.

embeddings
MiniLM-L6 · 384d
index
sqlite-vec KNN
scope
every meeting, ever
latency
< 80 ms
cache invalidation restsearching…esc
meetings
API redesign · vendor scoping
design_review · 2 weeks ago
91%
Cache strategy retro
retro · 4 weeks ago
71%
topics
REST over GraphQL · decision
decision · in API redesign · 2 weeks ago
84%
Webhook delivery · at-least-once
decision · in API redesign · 2 weeks ago
69%
transcripts
“…we’ll get further with REST and a CDN than fighting cache invalidation.”
API redesign · 2:22 · 2 weeks ago
79%
↑↓ navigate · open⌘K anywhere
§08 · the second brain at the table

Catches
what you forgot.

Magic2 listens. After about a minute it figures out what kind of meeting you’re having — a standup, a design review, a customer discovery, one of twenty-six templates — and starts watching for the items that type usually needs.

No more wrapping a standup having forgotten to ask about blockers. No more closing a design review without naming the decision. Magic2 nudges you, in the flow, before the conversation moves on.

Auto-detects
the meeting type in ~60 seconds
from twenty-six templates · re-evaluates as the conversation evolves
Watches for
the items that type usually needs
yesterday · today · blockers · asks · cross-team — the wee bits
Nudges you
before the meeting moves on
a soft prompt mid-flow · not a wall of red · easy to dismiss
auto-classifying
T+00:47
93%
confidence
type detected
Daily standup
Scrum · 15 min · 4 items
also considered
  • status_sync0.04
  • retro0.02
  • sprint_plan0.01
magic2 · meeting_type · full-lanere-evaluates every 180s
magic2 · coverage · daily standuplive · T+12:31
meeting typelocked
Daily standup
Scrum · 15 min · 3/4 covered
3/4
covered
suggested next
Yesterday's progress
What each person completed since the last standup
  • Yesterday's progress
    What each person completed since the last standup
  • Today's plan
    What each person will focus on today
  • Blockers
    Anything currently stopping progress
  • Sprint goal confidence
    Are we on track to hit the sprint goal?
last check · 4s ago · re-checks every 15s1 item still open
the nudge · in the flow of the meeting

A whisper, not a wall of red.

Coverage prompts appear as soft chips on the mind map, near where the conversation already is. Click to dismiss. Or just keep talking — Magic2 will check again next tick.

STANDUP · 2026-05-19Today's planBlockersSprint goal confidenceYesterday's progress?
suggested next
Yesterday’s progress
What each person completed since the last standup
daily standup · 1 of 4 opendismiss ×
26 · meeting templates known

From standups to legal review.

Each template knows the items it expects — the questions, the artifacts, the decisions a meeting of that shape should produce. Magic2 picks one as the conversation takes shape, and switches if it’s wrong.

standupdesign_review1:1retrosprint_planningcustomer_discoverysales_discoverysales_demohiring_loopperformance_reviewinvestor_pitchinvestor_updateboard_meetingall_handskickoffbrainstormrfc_reviewarchitecture_reviewbug_triagepostmortemvendor_pitchvendor_scopinglegal_reviewcontract_negotiationuser_interviewstatus_sync
anatomy · standup · template

What Magic2 listens for in a standup.

Pick standup because everyone has one and everyone forgets one of the bits. Here’s the full schema Magic2 watches for, every fifteen seconds, against the rolling transcript.

template · daily_standup · scrum · 15 min4 items watched · 15s tick
item 01
Yesterday's progress
What each person completed since the last standup
magic2 listens for
Past-tense actions by each speaker · ‘shipped’, ‘finished’, ‘reviewed’, ‘merged’
example match
“Shipped the cache fix and reviewed Lina’s PR.”
item 02
Today's plan
What each person will focus on today
magic2 listens for
Future-tense intent · ‘going to’, ‘today I’ll’, ‘picking up’
example match
“Going to spike the rate-limit redesign.”
item 03
Blockers
Anything currently stopping progress
magic2 listens for
‘Blocked on’, ‘stuck on’, ‘waiting on’, missing dependencies
example match
“Blocked on infra approval; can’t deploy until Tuesday.”
item 04
Sprint goal confidence
Are we on track to hit the sprint goal?
magic2 listens for
Confidence signals · ‘on track’, ‘slipping’, ‘at risk’, ‘ahead’
example match
“Still tracking for Friday but the auth piece is tight.”
type can be locked, or magic2 will keep re-classifyingv1.1 · custom templates →
preview
Magazine-quality PDF report
Magazine-quality PDF report
§09 · magazine pdf

Walk out with
a publication.

One click. Magic2 renders a print-ready PDF with proper IBM Plex typography — Sans for body, Mono for data, Serif for the editorial moments. Hero cover. Metric grid. One spread per artifact. Page breaks that respect the content.

Headline takeaway above the fold
At-a-glance metric grid (duration · word count · decisions · actions)
Each artifact gets a tailored spread
Legal memos + contract redlines carry a hard-coded disclaimer banner
Real page numbers, real footnotes, real fonts
§10 · built for

Different roles. Same conversation.

Magic2 doesn’t care what your job is — it cares what the conversation is. Six common shapes:

Product manager

Stop reconstructing decisions from Slack threads.

Magic2 emits the decision log with rationale and source quote per row. Plus action items, MVP spec, and risk register from the same conversation.

Founder

Investor updates that write themselves.

Run your update meeting, walk out with the memo. Thesis, market, team, traction, risks — structured the way LPs read it.

Engineer

Design reviews that survive past lunch.

Annotated mind map, RFC seed, action items with owners and dates. The architecture decision record that should exist.

Sales

Discovery calls that become CRM rows by 5pm.

Customer interview artifact with themes, pain points and verbatim quotes — pasted into Salesforce while the call is still warm.

Recruiter

Calibration meetings that become hiring briefs.

Role profile, must-haves, bar to clear, sample interview questions. From the same conversation that made the brief.

Legal

First-draft memo and redline while you're still on the call.

Issue / analysis / conclusion / recommendation, with the disclaimer banner you cannot remove and never want to remove.

§11 · local-first

The conversation
never leaves.

Audio streams to a WhisperLive instance on a GPU box you can see from your desk. Transcription and extraction run on a local LLM via Ollama. Vectors live in a SQLite file on your own disk. No third-party cloud touches the conversation unless you flip a switch.

audio
local · opus · disk
stt
whisperlive · your gpu
llm
ollama · your gpu
vectors
sqlite-vec · your disk
YOUR NETWORK · LANLaptopbrowser · audioGPU boxwhisperlive · ollamaqwen2.5:7b-instructSQLite+ sqlite-vecEXTERNAL CLOUD— not used —
§12 · under the hood

Boring tech. Exciting product.

Battle-tested parts in a quiet arrangement. Runs on one GPU box and a laptop. The whole thing fits on a single uvicorn worker.

STT
WhisperLive

WebSocket · base.en · 1s latency · reconnect-safe

LLM (fast)
qwen2.5:7b-instruct

Ollama · keep_alive=10m · imperative prompt · forgiving parser

LLM (heavy)
qwen3.5:9b

Post-stop artifacts · context-rich · latency tolerant

Embeddings
MiniLM-L6

384-dim sentence-transformers · CPU

Vectors
sqlite-vec

transcript_vecs + map_node_vecs · cosine KNN · embedded

Backend
FastAPI + uvicorn

Python 3.12 · async · WAL · WebSocket

Frontend
Next.js 16 + React 19

Tailwind 4 · react-force-graph-2d/3d · turbopack

Audio
MediaRecorder · Opus

2s timeslice · layout-scoped lifecycle

§13 · roadmap

Shipped. Shipping. On the horizon.

Shipping · v1.0 today
  • 01Live mind map
  • 02Eleven artifact kinds
  • 03Magazine PDF
  • 04Cross-meeting recall
  • 053D constellation
  • 06Cmd-K semantic search
  • 07Smart auto-naming
  • 08Continuation detection
Coming soon · v1.1—v1.3
  • 01Slack + Linear sync
  • 02DOCX / PPTX export
  • 03Shareable read-only URLs
  • 04Speaker diarization
  • 05Mobile / tablet UI
  • 06Quota throttling
On the horizon · v2.x
  • 01Devil's-advocate engine
  • 02Decision simulator
  • 03Fact-checker · web
  • 04Chat against this meeting
  • 05Self-host Docker bundle
  • 06Template marketplace
§14 · frequently asked

Things people always ask.

Q01How is this different from Otter / Granola / Fireflies?+
Those tools stop at the transcript or a summary. Magic2 starts there: a live mind map of the conversation, eleven separate post-meeting deliverables, cross-meeting semantic recall, a 3D constellation of every meeting you've ever had, and a magazine-quality PDF export. None of which any other meeting tool offers.
Q02Does my audio leave my network?+
No. The default is local-first: audio streams to a WhisperLive instance on a GPU box you control. Transcription and extraction run against a local Ollama. Vectors live in a SQLite file on your disk. Nothing leaves your network unless you explicitly opt into an external integration.
Q03What if the LLM hallucinates?+
Every artifact item carries a source-anchored quote and timestamp back to the transcript. If Magic2 claims a decision was made, you can click through to the exact moment in the audio. Legal memos and contract redlines carry a hard-coded AI-generated disclaimer that the LLM cannot remove.
Q04Do I need a GPU?+
For real-time use, yes — one consumer GPU is enough. We run on a Quadro P1000 for STT and a desktop GPU for LLM. CPU-only Whisper runs at 2.6–4× realtime and works fine for batch processing recorded meetings.
Q05Is it accurate?+
qwen2.5:7b-instruct with an imperative prompt and a concrete example produces clean, structured topic / decision / risk / action extraction. The forgiving JSON parser handles every reasonable variation. For artifact generation we use the larger qwen3.5:9b where quality matters more than latency.
Q06How long is a meeting before Magic2 falls over?+
We've stress-tested four-hour meetings without degradation. The mind map is force-directed with smart pruning of low-affinity edges. Extraction batches transcripts in rolling windows so context never blows out the LLM's working set.
Q07Can I export the data?+
Yes. Magazine PDF today; DOCX, PPTX, and direct Slack / Linear push are queued in the v1.1 milestone. SQLite is the source of truth — point any tool at it.
Q08Who's behind it?+
Solo developer. Built over a single, very-online quarter. Designed against the frustration of recording six meetings a day and walking out with no artifact except an inbox of stale Slack threads.
magic2 · v1.0 · shipping today

Stop taking notes.
Start having
the meeting.

There is genuinely nothing else like Magic2. You can keep using whatever you use today. Or you can try this and find out what it’s like to walk out of a meeting with eleven separate deliverables waiting for you.

local-first by defaultworks offlineno credit card