Linux · open source · experimental alpha

Remember what was decided.
Check where it was said.

Name: FalaTrace
Author: Caio Barbieri

FalaTrace records the calls you configure, transcribes with local processing by default and links summaries and notes to the moment in the recording. When your AI needs context, it retrieves only what it needs — with a timestamp and source you can check.

Install alpha.14 View code on GitHub

Free and MIT-licensed
No account, no hosted service
Nothing records just by installing

FalaTrace Studio Capture stopped

Recordings 3

Search recordings…

Weekly syncSep 29 · 10:00On this computer · Processed
Incident reviewSep 22 · 15:30On this computer · Processed
Q4 planningSep 15 · 09:00On this computer · Processed

Weekly sync

Sep 29 · 32 min · On this computer

Migration plan

MARFAN

Marina: Let’s push it to sprint 42 — rollback isn’t covered yet.

12:48 / 32:10

TranscriptSummarySpeakersAI context

AI summary · check the source

Decisions

Postpone the database migration to sprint 42.12:48
Keep Friday’s deploy behind a feature flag.18:31

Action items

Ana reviews the rollback plan by Friday.14:02

1 reference with approximate timing

IllustrationFictional data; the Studio is available in English and Portuguese. See the real Studio screens below.

Automatic detection in: Slack · Zoom · Chromium · Firefox · Zen · Helium and more
Processes locally with: Whisper.cpp · Ollama
Or, if you choose: OpenAI · Gemini
Context for agents like: Claude Code · Codex · Gemini CLI

How it works

From call to context, without losing the source.

Every step stays under your control. You choose when to record, where to process and what to share.

01Record New in alpha.14

Records the calls you choose.

Turn on rules to record when Slack, Zoom, Chromium, Firefox, Zen, Helium and other apps use the microphone or camera — or start capture manually. Recording state is always visible.

Detection stays off until you enable it; the initial mode only notifies.
“Pause new recordings” and “Stop” are separate actions.
An assistant guides the first use, and Settings covers recording, capture, AI, API keys, models and services.
Microphone and system audio on separate tracks; screen via OBS or GPU Screen Recorder.

REC 02:08 Slack · microphone and camera in use Stop

Automatic recording

SlackOn
ZenOn
HeliumOff

Pause new recordingsThe current capture continues until you stop it.

IllustrationFictional data.

02Check

Summaries that point to the source.

Follow each decision back to the transcript segment and the moment in the video. Approximate timing and uncertain references are flagged for review.

References carry the segment, the timestamp and the recording’s SHA-256 hash.
Transcription with Whisper.cpp on your computer, or OpenAI and Gemini if you configure them.
Validation proves the reference exists — not that the model interpreted it correctly.

Summary · decision

Postpone the database migration to sprint 42.

s000214clear reference

Transcript · segment s000214

“Let’s push it to sprint 42 — rollback isn’t covered yet.”

12:48.4 → 12:55.1Marina · segment timing

Original recording

weekly-sync.mkv

sha256 3f9c…e1a0checked before and after processing

IllustrationFictional data.

03Review

Review without erasing the original.

Fix segments and speaker names in the Studio, generate a new summary with your consent and export exactly the version you checked.

History with undo; the original media, transcript and speakers are preserved.
A new local summary runs only after you see the model, destination and limits — and confirm.
Export to JSON, Markdown, SRT or WebVTT from the preview you reviewed.

12:48MarinaReviewedUndo

Let’s push it to ~~sprint forty two~~ sprint 42 — rollback isn’t covered yet.

Original preserved · revision 3

Export revision 3

JSONMarkdownSRTWebVTT

Reviewed preview · segments with edited words export as JSON or Markdown

IllustrationFictional data.

For your AI

Context with a source for your agents.

Agents running on your computer — like Claude Code, Codex and Gemini CLI — query your recordings through the CLI or a local MCP server. They receive bounded excerpts with a timestamp, source and hash, and only within what you authorized.

falatrace · CLI

$ falatrace context search "database migration" --limit 3
{
  "items": [{
    "jobId": "7c1e…b42f",
    "title": "Weekly sync",
    "date": "2026-09-29",
    "excerpt": "…let’s push it to sprint 42 — rollback…"
  }],
  "trust": "untrusted-meeting-data"
}

$ falatrace context meeting 7c1e…b42f --query "migration"
{
  "excerpts": [{
    "timestamps": { "start": 768.4, "end": 775.1 },
    "timing": "segment",
    "segmentIds": ["s000214"],
    "text": "Let’s push it to sprint 42 — rollback…"
  }],
  "budget": { "maxCharacters": 4096, "truncated": false },
  "trust": "untrusted-meeting-data"
}

ExampleAbbreviated output with fictional data; the fields are the CLI’s own.

You

Why did the database migration move to sprint 42?

Your agent

In the Sep 29 weekly sync, Marina asked to postpone it because rollback wasn’t covered yet 12:48. Ana will review the plan by Friday 14:02.

IllustrationFictional conversation. Answer quality depends on the model and has not been validated.

Ready-made skill
Teaches the agent to search, cite the source and respect limits — with no permission to record, upload or delete.
Access you control
Grants per recording and recipient, with pause, revocation and a budget that doesn’t reset on restart.
Data, not orders
Meeting content arrives labeled as untrusted data, never as instructions for the agent.

No agent is connected automatically: you configure each client. Agents backed by cloud models may send what they receive to their provider.

Studio

A desktop app to review everything in one place.

Library, video, transcript, summary, speakers and AI context. These are real captures of the Qt interface with synthetic test data — nobody was recorded. They show the Portuguese interface; since alpha.14 the Studio is also available in English.

The Studio in Portuguese, light theme: list of recordings on the left, video area in the center and transcript on the right, with a synthetic sample recording. — Real UI · alpha.5Library, video and transcript in the light theme.

The Studio in the dark theme with the “Antes de iniciar a gravação” (before recording) dialog, asking to confirm permission from everyone involved. — Real UI · alpha.1Confirmation before recording, in the dark theme.

The Studio’s “Acesso para IA” (AI access) dialog with the local extraction and analysis budget and a preview of a synthetic color-bar frame. — Real UI · alpha.5AI access: grant, budget and frame preview.

These captures come from earlier alphas; the Studio was redesigned in alpha.12, gained Settings in alpha.13 and a first-use assistant in alpha.14, so it looks different today. Click an image to enlarge it.

Privacy and control

Your data stays where you decide.

There’s no account and no hosted service: processing happens on your computer or at the destinations you configure.

Local by default
A new configuration uses Whisper.cpp and Ollama on your computer, with automatic processing turned off. You install the models.
Sending is your choice
OpenAI, Gemini, remote workers and backups only run if you configure them — and they can transmit content. Check destination and cost first.
Consent at every step
Recording, generating a Studio summary and releasing frames to an agent are separate authorizations, with limits that persist.
Deletion with a clear scope
Local removal invalidates derived memory before touching files. Remote copies, or data already accepted by a provider, are not erased by it.

Honest status

Used in the author’s routine. Still an alpha.

Caio Barbieri created FalaTrace and continuously uses its recording and context-recovery core. That report is not a reliability measurement — here is what has been verified and what remains to be demonstrated.

Verified offline

931 tests with no failures in alpha.14, using synthetic data.
Locked installation, configuration that doesn’t overwrite yours and uninstall that preserves data.
Eight Studio journeys with real QML in offscreen Qt, 323 checks, running in CI (alpha.14).

Not yet demonstrated

Detection of the new apps in real calls and capture in other Linux environments.
Real-world quality of transcription, speakers, summaries and visual analysis with installed models.
Screen readers, varied display scaling, installing on a machine without a package cache, binaries and other platforms.

Compatibility in alpha.14
Capability	Status
Automatic detection in Slack, Zoom, Teams for Linux, Chromium, Chrome, Brave, Edge, Vivaldi, Opera, Firefox, Zen and HeliumVia local microphone or camera signals; it doesn’t identify the calling service. The apps new in alpha.12 were tested with synthetic data.	Available
Discord, Signal, Telegram and ElementOff by default because they also carry personal calls; turn each one on in the configuration.	Optional
Full settings in the StudioEvery option, API keys in a private file, verified model downloads and services, with diagnostics.	Available
Manual recording (FFmpeg audio, OBS, GPU Screen Recorder)Depends on the backend and desktop.	Available
Transcription with local Whisper.cpp, OpenAI or Gemini	Available
Summaries with local Ollama or OpenAIModel quality not validated.	Available
Desktop Studio (Qt)English and Portuguese interface.	Available
Windows and macOS	Out of scope

Start from source

Install alpha.14 in a few commands.

The alpha ships as source code for Linux. None of these commands starts a recording: automation is enabled separately, after you review rules and consent.

You’ll need

Linux with Bun 1.4 and FFmpeg
Optional: Whisper.cpp and Ollama for local processing
Optional: Qt 6.10 with MpvQt for the Studio, from system packages or bun run desktop:setup

Terminal

# 1. Get the code and pin alpha.14
git clone https://github.com/caiolombello/falatrace.git
cd falatrace
git switch --detach v0.2.0-alpha.14

# 2. Install the CLI into ~/.local/bin
bun install --frozen-lockfile --ignore-scripts
make install-cli INSTALL_PREFIX="$HOME/.local"

# 3. Optional: install the Studio; on first launch an assistant sets everything up
make install-studio

# 4. Without the Studio: create a config and check your setup (records nothing)
falatrace init
falatrace record doctor

init only creates a config when none exists; existing configurations are preserved. record doctor only inspects the backends. make install-studio uses no sudo and downloads nothing.

FAQ

Frequently asked questions

Didn’t find your answer? Open an issue — without pasting configuration or private data.

Does FalaTrace start recording on its own after installation?

No. Installing and running falatrace init doesn’t start any capture. Automatic detection ships turned off; when enabled, its initial mode only notifies. Recording automatically requires an explicit configuration on your part.

Which calling apps does it recognize?

Slack, Zoom (native app), Teams for Linux, Chromium, Chrome, Brave, Edge, Vivaldi, Opera, Firefox, Zen and Helium, via local microphone or camera signals — plus Zoom web, Meet and Teams web inside those browsers. Discord, Signal, Telegram and Element can be turned on individually. The apps new in alpha.12 were tested with synthetic data, not yet in real calls of each app. That signal doesn’t identify the service (Meet, Teams, Zoom on the web) or the tab. For other apps, use manual recording.

Do my recordings go to the cloud?

Only if you configure it. A new configuration processes locally with Whisper.cpp and Ollama. Providers like OpenAI and Gemini, remote workers and backups are optional and can transmit content; the Studio shows destination and limits before asking for your consent to summarize.

How much does it cost?

FalaTrace is free and open source (MIT). External providers bill separately, and a chat subscription (ChatGPT, Claude, Gemini) doesn’t include API usage.

Does it work with ChatGPT, Claude or Gemini in the browser?

Not directly. The integration is for agents that run on your computer and use the CLI or the local MCP server, like Claude Code, Codex and Gemini CLI. Hosted websites can’t reach local tools through it.

Can I trust the summaries?

Always check. Summaries point to transcript segments and flag approximate timing and uncertain references. Validation proves the reference exists — not that the model interpreted it correctly.

What about consent from other people on the call?

That’s the responsibility of whoever records. The Studio asks for confirmation before starting a capture, and automatic detection doesn’t establish consent. Follow the laws and policies that apply to you.

Is there a Windows or macOS version?

No. This alpha is Linux-only and distributed as source code.

Is the interface available in English?

Yes. Since alpha.14 the Studio follows the system language, English or Portuguese, and can be switched in Settings. The CLI, the agent skill and the technical documentation are in English.

Stop hunting for what was already said.

Install alpha.14, configure it at your own pace and give your AI context with a source.

Install alpha.14 View the release on GitHub

Remember what was decided.Check where it was said.

From call to context, without losing the source.

Records the calls you choose.

Summaries that point to the source.

Review without erasing the original.

Context with a source for your agents.

Ready-made skill

Access you control

Data, not orders

A desktop app to review everything in one place.

Your data stays where you decide.

Local by default

Sending is your choice

Consent at every step

Deletion with a clear scope

Used in the author’s routine. Still an alpha.

Verified offline

Not yet demonstrated

Install alpha.14 in a few commands.

You’ll need

Frequently asked questions

Stop hunting for what was already said.

Remember what was decided.
Check where it was said.