Natively AI Assistant
A local-first desktop copilot for live meetings and interviews that transcribes in real time, drafts answers and notes, and keeps your data on your own machine.
- Stars
- ★ 3k
- Last updated
- 1d ago
- License
- NOASSERTION
- Primary language
- TypeScript
- FA score
- 43/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI API · Claude API
- Cost
- Free tier plus a paid hosted plan
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A job seeker in a technical or behavioral interview who needs concise real-time answer guidance and code explanations
- Not a fit if
- Users who need Linux desktop support (currently unavailable)
- People wanting a zero-config, all-in-one cloud service with no API keys
- Candidates trying to evade dedicated proctoring software like Pearson VUE or ProctorU
- Source review
- 43/100 · Major gaps 1 safety controls not found
What does this agent do, and when should you use it?
Natively is an Electron/React desktop app for macOS and Windows with a Rust native audio module, positioned as a real-time meeting assistant, interview copilot and note taker. It captures system audio and microphone input on separate channels, transcribes through on-device Whisper/Moonshine/Parakeet models or cloud STT providers such as Soniox, Deepgram and Google, and generates live answers and meeting notes using your own LLM keys (Gemini, OpenAI, Anthropic, Groq, Ollama and others). Data is stored locally, with SQLite and sqlite-vec powering a local RAG index for searching past meetings semantically. The app adds an always-on-top overlay, screenshot OCR, nine built-in persona modes and a toggleable stealth mode. It is free for personal, educational and non-commercial use, with a paid Pro edition and a hosted Natively API tier on top.
A Rust native module captures system audio and microphone input at sub-500ms latency, applies sample-rate detection and WebRTC VAD silence filtering, then sends audio to a speech engine (local Whisper-family models, Apple Speech on macOS 26+, or a cloud STT provider). Transcribed text flows into a rolling context window alongside reference files (PDF/DOCX/TXT), custom notes and job descriptions before being routed to your configured model endpoint. The overlay can screenshot the screen and OCR it, handing LeetCode or HackerRank problems to GPT, Claude, Gemini or Ollama; Visual Answers draws flowcharts, sequence diagrams and charts inline while the answer streams. After a meeting, a local SQLite + sqlite-vec index enables semantic search over history, and a dashboard exports notes as Markdown, JSON or text. Skills are local SKILL.md persona files invoked from the overlay chat with / or $.
- A job seeker in a technical or behavioral interview who needs concise real-time answer guidance and code explanations
- A sales or recruiting professional who wants structured meeting notes and a follow-up email draft after each call
- A student recording a lecture or seminar who later wants to search the transcript by meaning
- A developer facing a LeetCode or HackerRank problem who wants a solution and complexity analysis quickly
- A remote team that wants a searchable local meeting knowledge base without uploading recordings to a third party
How do you install or deploy this agent?
Install on macOS via Homebrew, or download a build from the Releases page:
brew install --cask Natively-AI-assistant/tap/nativelyIf macOS reports the app is damaged, run:
xattr -cr /Applications/Natively.appBuilding from source requires Node.js 22.13+, Git, and Rust (plus Xcode 26+ for the Apple Speech helper):
git clone https://github.com/Natively-AI-assistant/natively-cluely-ai-assistant.git
cd natively-cluely-ai-assistant
npm install
npm run build:nativeCreate a .env file in the project root with your keys, for example:
GEMINI_API_KEY=your_key
OPENAI_API_KEY=your_key
CLAUDE_API_KEY=your_key
USE_OLLAMA=true
OLLAMA_MODEL=llama3.2
OLLAMA_URL=http://localhost:11434
DEFAULT_MODEL=gemini-3.1-flash-lite-previewHow do you use this agent?
Run the app in development mode:
npm startBuild a production package:
npm run distOn first run, configure at least one speech provider (Apple Speech needs no key on macOS 26+, or supply a Soniox, Deepgram or Google STT key) and one LLM provider (a cloud API key or a local Ollama instance). Auto Answer (Beta) is off by default and lives under Settings → General; embeddings and rerankers are chosen under Settings → Retrieval; stealth and process disguise are toggled in settings. Day to day, Cmd+B / Ctrl+B shows or hides the overlay, Cmd+K / Ctrl+K opens spotlight search, and a screenshot shortcut captures on-screen problems.
What are this agent's strengths and limitations?
- Native audio capture is documented at sub-500ms latency with separate system-audio and microphone channels, which avoids room-noise contamination
- BYOK across major providers plus local Ollama enables a fully offline transcript-and-inference path, with data kept in a local SQLite + sqlite-vec index
- Nine built-in modes (technical interview, sales, recruiting, lecture, seminar and more) plus SKILL.md custom personas
- One app covers live transcription, screenshot OCR, meeting notes, history search and export instead of stitching tools together
- Free for personal, educational and non-commercial use, with stealth mode included rather than sold as a paid tier
- No Linux support; the README explicitly asks for maintainers to help add it
- You must supply LLM and speech provider keys or install Ollama, which is more setup than an all-in-one cloud tool
- Several features (Hindsight long-term memory, Reference Files, speaker diarization, custom personas) require a paid Pro license
- Stealth is documented as best-effort and does not cover dedicated proctoring software such as Pearson VUE or ProctorU
- Feature-usage telemetry is on by default, though it never contains spoken, viewed or typed content and can be disabled
How does this agent compare with similar options?
The README explicitly positions Natively against Cluely, Final Round AI, LockedIn AI, Interview Coder, Pluely and Parakeet AI. Against Cluely it emphasizes local storage, BYOK and free stealth; against Final Round AI ($149/mo) and LockedIn AI ($55–70/mo) it argues for zero subscription and any-model flexibility; against Interview Coder it claims parity on process-level disguise while covering behavioral and system-design rounds for free. It credits Pluely as a lightweight Tauri alternative with Linux support, while arguing Natively is more complete on RAG, meeting history and dual-channel audio.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Natively AI Assistant This agent | 43 · Major gaps | Desktop appFreemium | ★ 3k | 1d ago | TypeScript | OpenAI API · Claude API |
| Grux for Mac | 67 · Some gaps | Desktop appFree | ★ 30 | 4d ago | Swift | Claude API |
| PyGPT Desktop Assistant | 49 · Major gaps | Desktop appFree + model costs | ★ 2k | 15d ago | Python | OpenAI API · Claude API |
| Rowboat | 43 · Major gaps | Desktop appFree + model costs | ★ 18k | today | TypeScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows a detailed privacy-boundary architecture doc (usage-ledger-architecture explicitly excludes prompts, transcripts, screenshots, keys; stores only SHA-256 fingerprints), so data_flow_transparency scores 2. Deductions are concrete: README claims 'no data breaches' and 'runs locally', yet docs reveal a server-side usage ledger, Dodo webhooks, and Axiom/PostHog/Sentry telemetry, and the progress doc admits usage_events remains UPDATE/DELETE-able by service_role (O3 unfixed), so least_privilege is 1. User confirmation: Auto Answer is off by default and telemetry can be disabled, but no per-action confirmation for recording/screen capture/stealth is evidenced, so 1. Sensitive data: docs promise no content storage, but 'Custom Context can hold credentials' plus stealth/process-disguise features carry high abuse risk, and no evidence of encrypted key storage, so 1. Dependency security: only dependency names listed, no lockfile audit or vulnerability scanning evidence, so 0. External effects: process disguise, hidden Dock/taskbar, and analytics suppression are significant system-level side effects justified only as 'best-effort', so 1. Rollback: kill switches and per-call flag reads exist, but no version rollback or migration-reversal guidance, so 1. Source attribution: LICENSE has attribution clauses, but README's competitor comparisons and marketing claims lack verifiable sourcing, so 1.
Self-consistency shows clear contradictions: README says 'source-available, free' while LICENSE forbids commercial use and states it is not open source; the progress doc says 'NOTHING OUTSTANDING' yet lists O1/O2/O3 as open; 'Nothing is committed' coexists with 'migrations applied', so 1. Dependency availability: only fastify, supabase etc. named, no version pinning or install verification evidence, so 1. Failure messages: docs record specific failure texts and fixes (source-authority refusal wording, harness needing NATIVELY_TRACE_LONGCTX), showing error handling, so 2.
Audience and scenarios: README covers interviews, meetings, sales, recruiting, lectures across 9 modes, so 2. Capability boundaries: explicitly states stealth is best-effort, cannot defeat dedicated proctoring, Windows untested — honest boundaries, so 2. Trigger precision: Auto Answer is Beta and 'decides whether an answer is wanted' with no precision evidence, so 1. Environment fit: states macOS 12+/Windows 10/11, Apple Silicon/Intel, offline Ollama, so 2.
Information architecture: README is extremely long and marketing-heavy, burying install and limitations under competitor tables, so 1. Install notes: Homebrew cask, download links, and system requirements present, so 2. Naming stability: repo name vs star badge point to two different owners (evinjohnn vs Natively-AI-assistant), inconsistent attribution, so 1. Examples and FAQ: demo GIF and some anchor Q&A but no structured FAQ, so 1. Known limitations: stealth not guaranteed, Windows untested, no Linux support stated, so 2. License: LICENSE is complete and clear (Personal Use Source License v1.0), so 2. Versioning/changelog: only version 2.9 mentioned, no CHANGELOG, so 1. Maintenance responsibility: LICENSE names Natively AI Private Limited, but publisher is unverified and docs show staged uncommitted work, so 1.
Output usability: real-time transcription, visual answers, meeting notes, export and dashboard give rich output forms, so 2. Marginal value: core functionality heavily overlaps Cluely et al.; differentiation is 'free + local + stealth', but stealth carries ethical/compliance risk, so 1. Cost benefit: BYOK and local Ollama can cut cost, but README's '$0' conflicts with LICENSE commercial restrictions and actual API token costs, so 1.
Claim traceability: README asserts 83,000-user breach, <500ms latency, 300k+ downloads, 9,000+ users with no cited sources, so 1. Cross-source corroboration: docs contain self-corrections (CORRECTIONS section) allowing partial cross-checking, but README marketing claims conflict with docs, so 1. Fact/inference separation: docs distinguish 'server-observed' vs 'client_reported' well, but README presents inferential comparisons as fact, so 1.
- Not found in source: dependency securityPin versions and run a dependency audit (npm audit, pip-audit) before installing; prefer running it in a container.
- README marketing claims (competitor breaches, latency, user counts, downloads) lack verifiable sources and should not be taken as fact.
- LICENSE is a Personal Use Source License that forbids commercial use, in tension with README's 'free/open' framing; enterprise use requires separate commercial licensing.
- Stealth/process-disguise/hidden Dock and taskbar features may violate interview or meeting platform terms, raising compliance and ethical risk.
- Docs admit usage_events remains UPDATE/DELETE-able by service_role (O3 unfixed), so the append-only ledger guarantee does not yet hold.
- Server-side telemetry (Axiom/PostHog/Sentry) and a usage ledger exist, not fully consistent with 'fully local' claims; verify against PRIVACY.md.
- Repo name and star badge point to different owners (evinjohnn vs Natively-AI-assistant); publisher identity is unverified.
- Docs show staged uncommitted work and a server.js conflict history, raising build reproducibility concerns.