Second Brain Personal AI Assistant
Turns an LLM into a personal operating system on your own machine: local file intelligence, workflow automation, and multi-platform messaging so tasks complete themselves.
The README describes high-risk capabilities (local file indexing, shell execution, Telegram, Gmail, cron) but the provided sources show no permission boundaries, confirmation flows, data-flow documentation, or sensitive-data handling; a sandbox appears only as a test fixture with unverified constraints. requirements.txt is minimal with no pinning, lockfile, or dependency-security review — clear deductions. MIT LICENSE and author attribution (Henry Daum) are present, earning basic attribution credit only. No rollback or uninstall mechanism is evidenced.
No dependency is version-pinned (no == constraints), so upstream breakage directly affects installs: dependency_availability is 0. Comments describe ~2200 tests with pytest-xdist parallelism, showing self-consistent test infrastructure (1). README offers some troubleshooting (401/blank screen). Core code is absent from the sample, so most reliability claims cannot be verified.
The README clearly targets a personal local-machine assistant across Windows/macOS/Linux, Docker, Telegram, and Web UI — audience and environment fit earn 2. Capability boundaries (what not to delegate) and trigger precision (event-driven task conditions) are barely documented, scored 1 each.
Install documentation is exceptionally thorough (venv, Docker, UI token setup, per-platform tables): 3. LICENSE is complete MIT: 3. Data-directory layout, /locations, /packages give a reasonable information architecture: 2. But there is no CHANGELOG or version number (0), known limitations are one throwaway sentence, naming stability and maintenance responsibility rest solely on one individual author with no governance or update commitment (1 each). Examples are mostly infographics and external links, not verifiable in-repo.
The README promises strong functionality but the sample contains no core implementation, so output usability and marginal value rest on description alone (1 each). Benchmark superiority over OpenClaw/Hermes is shown only as a screenshot from a self-built harness by the author; the cost-benefit case is under-supported.
The benchmark is self-built and self-run, so cross-source corroboration is 0; claims trace to the author's own evals repos (1). The README largely separates factual claims from the author's 'Philosophy' musing, so fact/inference separation earns 1. No execution was performed; confidence stays low.
- This is a static source review; no code was executed, and core implementation files are absent from the sample, so most functional claims are unverified.
- Dependencies are unpinned with no lockfile; installs may break on upstream changes — review and pin requirements yourself before deploying.
- The agent ships with high-risk capabilities (local file indexing, shell execution, Telegram/Gmail access) from an unverified publisher; trial it in isolation (Docker or a dedicated account) and be cautious granting sync_directories and messaging credentials.
- The README's benchmark is self-built and self-run, presented only as a screenshot; treat it as marketing, not independent evidence.
What does this agent do, and when should you use it?
Second Brain is an open-source agentic framework (GitHub: henrydaum/second-brain, MIT licensed) designed to run as an AI operating system on your own computer. Its architecture spans attachment parsing, file parsing, an LLM loop, path-driven and event-driven tasks, a conversation runtime, frontends, commands, plugins, and a sandbox/SDK, combining RAG, tool calls, and cron jobs to get real work done. It runs as a REPL by default, with optional Telegram bot and a separate second-brain-ui React web frontend built on assistant-ui. Installation requires Python 3.11+ and git, with an optional Docker deployment; a /setup wizard configures the essentials bundle, a model provider via API key (Atlas Cloud is the sponsored fast path, but any provider works), and Telegram/Web UI in one pass. Capabilities extend through /packages, including bundle_knowledgebase (indexing and searching PDFs, Office files, images, audio, video, spreadsheets, and archives), bundle_memory (self-maintaining durable memory), and bundle_gmail.
Run python main.py to enter the REPL, then /setup to install the essentials bundle (LLM backend, file read/edit/search, shell and script running, SQL, web search, subagents, and the Telegram frontend) and connect a model by pasting an API key. It parses attachments and local files and completes tasks in an LLM loop with tool calls; /config sets sync_directories to index chosen folders (OCR, transcription, embeddings, and three search tools via bundle_knowledgebase). It supports path-driven tasks, event-driven tasks, and cronjob scheduling; reads, sends, replies, and labels Gmail (bundle_gmail); and communicates over multiple modalities via Telegram or the HTTP frontend (frontend_http plus second-brain-ui, a React app on assistant-ui). Key commands include /packages, /config, /llm, /conversations, /commands, and /locations.
- Individuals who want a searchable knowledge base built from PDFs, Office documents, and audio/video scattered across their machine
- Users who want to chat with their assistant from their phone with push notifications via Telegram
- Productivity users automating repetitive work with cron jobs and event-driven tasks
- Developers who prefer a terminal REPL where the model can read/edit files and run scripts on their machine
- Office workers who need Gmail automatically read, sent, replied, and labeled
- Users running the assistant long-term on a NAS or server via Docker
What are this agent's strengths and limitations?
- Local-first architecture: config, database, packages, and workspace stay on your own machine, suiting privacy-conscious personal assistant use
- Capabilities extend on demand via bundles (bundle_knowledgebase, bundle_memory, bundle_gmail), keeping fresh installs deliberately small
- Multiple frontends: REPL, Telegram bot, and a ChatGPT-style Web UI you can add to your phone's home screen
- Documented benchmark performance: outperforms OpenClaw and Hermes on harness-bench across 106 sandboxed offline agent tasks, with the eval framework and results publicly available
- The knowledge-base bundle is a large download, and the first folder sync floods you with task messages
- The Web UI needs a separate frontend repo, Node, and manual token configuration; production builds ship no API token and require a reverse proxy, documented only for macOS deployment
- Docker containers cannot see your host machine's files, weakening the core local-file-intelligence value; the README itself recommends local installation for daily use
- Depends on an external LLM API key (Atlas Cloud recommended, any provider works), so cost and model quality depend on your chosen provider; the README offers no detailed per-provider adaptation notes
How do you install or deploy this agent?
- Install Python 3.11+ and git. 2. git clone https://github.com/henrydaum/second-brain && cd second-brain && python -m venv .venv; activate with .venv\Scripts\activate (Windows) or source .venv/bin/activate (macOS/Linux). 3. pip install -r requirements.txt && python main.py. Or Docker: docker build -t second-brain . && docker run --rm -it --init -v sb-data:/data second-brain. 4. In the REPL run /setup: install the essentials bundle, paste an API key to connect a model (Atlas Cloud is the sponsored fast path; any provider works), and optionally configure Telegram (bot token from @BotFather plus your user ID from @userinfobot) and the Web UI. Optional UI: also requires Node.js 20.19+ or 22.12+; clone second-brain-ui, npm install, copy .env.example to .env.local, set VITE_SB_TOKEN, then npm run dev and open http://localhost:5173.
How do you use this agent?
After startup you land in the REPL; run /setup first. Use /packages install to add capability bundles (also available in Web UI Settings), /config to set sync_directories for folders to index (expect a flood of task messages during the first sync — it stops when finished), /llm to add or switch models, /conversations to browse chats, /commands to list everything, and /locations to find your data directory (Windows: %LOCALAPPDATA%\Second Brain\; macOS: ~/Library/Application Support/Second Brain/; Linux: ~/.local/share/Second Brain/). The Web UI token is set in /config under secret_http_token, followed by /restart; for phone access run the dev server with --host and optionally reach it remotely via Tailscale.
How does this agent compare with similar options?
The author benchmarks Second Brain against OpenClaw and Hermes on harness-bench (106 sandboxed offline agent tasks), where Second Brain scores higher; the evaluation framework is published at second-brain-evals and full results at second-brain-eval-results, though the author notes a few small caveats.