Skip to content
AsumaCodes
Menu

Murmur

Say an idea out loud. Murmur researches the market, writes the PRD, the brand kit and the engineering brief, then builds the Jira board and Confluence space in your own Atlassian account.

Film: a real Murmur run, ShipReel. The 49-second memo against the 7-minute run, the six screens it produced, where the time goes, and the measured range across four public runs.

One public run, ShipReel, from a 49 s memo. Every screen is a real capture and every number is read from the run file.

Try Murmur (opens in a new tab)

Record the idea in the app, up to ten minutes. Press Run. Murmur works through four stages and hands back eight artifacts, and the Jira board and Confluence space are real, in your own Atlassian account.

It is a foundation, not a finished app. It writes the plan and sets up the board; it doesn't write the product's code.

Status
Early access
Opens
1 October 2026
Built
April to September 2026
Stack
Next.js, Supabase, n8n, AssemblyAI, Exa, Claude, Atlassian APIs, Dodo Payments
Role
Product, design and engineering. Built with Claude Code.
Links
trymurmur.studio (opens in a new tab)

What comes out of one memo

Eight artifacts per run

  1. 1 of 6

    Competitor map

    Exa searches the market. Each competitor gets its positioning, pricing model and overlap with the idea. ShipReel's run found 5.

    ShipReel's competitor map in Murmur: five competitors found from a 49-second memo.

  2. 2 of 6

    PRD

    The problem, the users, success metrics and features sorted must, should, could and won't.

    ShipReel's PRD in Murmur: problem, users and features sorted must, should, could and won't.

  3. 3 of 6

    Brand kit

    A tagline, a palette with hex values, type and brand values, downloadable as a zip. ShipReel's tagline: “Repo in, launch video out.”

    ShipReel's brand kit in Murmur: the tagline, seven colours with their hex values, and type.

  4. 4 of 6

    Engineering brief

    Architecture, data models, a database schema and the engineering tasks, ready to plan against.

    ShipReel's engineering brief in Murmur: architecture, data models and the build plan.

  5. 5 of 6

    Jira board

    Epics and stories created in your Jira through Atlassian's API. ShipReel: 11 epics, 33 stories.

    ShipReel's Jira backlog, created by Murmur: 33 stories under 11 epics.

  6. 6 of 6

    Confluence space

    Six pages: Product, Brand and Identity, Engineering, Competitor Analysis, Roadmap and Research Notes.

    ShipReel's Confluence roadmap page, created by Murmur, one of six pages in its space.

Four public runs, measured

These are the runs published on Murmur's site. Each number is read from its run file, and the build that draws this chart checks that every run's epics add up to its stories.
7 m 02 s to 9 m 22 s

  1. Creature Clash: Duel Loop Prototype

    Mobile creature-battler game prototype

    Epics
    12
    Stories
    47
    Competitors
    2
    2 m 56 s memo
    9 m 22 s run

  2. ShipReel

    Launch-video SaaS for founders

    Epics
    11
    Stories
    33
    Competitors
    5
    49 s memo
    7 m 02 s run

  3. Aavaas

    AI home-design & floor-plan tool

    Epics
    11
    Stories
    39
    Competitors
    6
    1 m 57 s memo
    7 m 42 s run

  4. Fin Matter

    Personal AI finance assistant

    Epics
    11
    Stories
    41
    Competitors
    7
    2 m 57 s memo
    7 m 31 s run
Thin line: the voice memo. Heavy line: the run, from start to results delivered. One scale for both.Source: Murmur's public example runs (run files in the Murmur repo)

Where the time goes

10 m 59 s, one recorded run

The last stage takes the longest. It writes the engineering brief and the roadmap, then creates every epic, story and Confluence page through Atlassian's API.

This run took 10 m 59 s, longer than any of the four public ones, which is why it's labelled on its own.

  1. Researching the market90 s
  2. Writing the PRD111 s
  3. Designing the brand26 s
  4. Building your board422 s
Total 10 m 59 s, from pressing Run to the results. A founder's run, recorded 26 September 2026; not one of the four public examples. Timed from its screen recording, from pressing Run to the results.Source: Timestamps read off the run's screen recording, each one checked

How it works

Four stages, eight artifacts

  1. Record

    Up to ten minutes in the app, on a phone or a laptop. The memo is transcribed as soon as recording stops.

    Next.js app, AssemblyAI

  2. Checks before it runs

    A transcript, a connected Atlassian account and enough balance. If the balance check fails, the run doesn't start.

    Supabase

  3. A signed handoff

    The app sends the audio to the pipeline with a signature over the run, a timestamp and a hash of the audio, so the pipeline only runs what the app sent.

    HMAC-SHA256, to the Bridge on n8n

  4. Research

    Competitors and the market around the idea.

    Exa

  5. PRD, brand kit, engineering brief

    Written one stage at a time from the transcript and the research.

    Claude

  6. The board

    Epics, stories and a six-page Confluence space, created in the user's own account. Access tokens are stored encrypted.

    Atlassian OAuth, AES-256-GCM

  7. Live progress

    Each stage reports back as it finishes, so the app shows the run moving in real time.

    Supabase Realtime

Decisions

Numbers where there are numbers

  1. Screens captured behind a privacy gate

    The screenshots here come from a script, not a human with a screenshot key. It hides account menus, avatars and bylines before the page paints, crops the Atlassian chrome, scans each screen's text for emails, keys and tokens, and drops any screen with a hit.

    24 screens from 4 runs

    Captured for Murmur's site, each one text-scannedSource: Murmur's capture script and its committed captures

  2. A judge that can't inflate a score

    Runs are scored by deterministic checks first. An LLM judge can only lower that score, never raise it: the final score is the check score times (0.5 + 0.5 × the judge's rating), so a perfect rating leaves it unchanged.

    19 deterministic checks

    Across the PRD, research, engineering, Jira, operations and one cross-checkSource: Murmur's quality-judge code

  3. Numbers in films are recomputed

    Before a Murmur film renders, a guard recomputes every number on screen from the run it shows. If one doesn't match, nothing renders.

  4. A page that doesn't jump

    Jumping straight to the run explorer used to shift the whole layout. Replays that swapped content for skeletons were replaced with transforms and colour changes.

    0.43 to 0.000

    Layout shift on the run explorer, before and afterSource: Murmur commit 7d01139

  5. Fail closed on money

    If the balance check itself errors, the run is refused rather than let through. A failed run never uses up an idea.

  6. Tokens encrypted at rest

    Atlassian access tokens are encrypted before they're stored and refreshed with rotation a minute before they expire.