Corporations have structured data about you.
Now you have one too.
They know you better than you know yourself.
Corporations have spent decades building structured datasets about every person on the internet. You have none of that about yourself. As AI gets more capable, this gap becomes a chasm.
Data asymmetry
Corporations invest billions structuring data about you — you don't even have access to most of it. The agent that knows you best isn't yours.
Context-blind AI
AI agents are powerful but useless without personal context. Your data is scattered across dozens of siloed apps. Your agent knows none of it.
No guided path
Learning AI has no guided path. You either figure it out alone, get overwhelmed by options, or never start at all.
The gap between what AI could know about you and what it actually knows — that's the problem we're closing.
An agent that actually knows you.
marsbot learns your context, builds a structured dataset, and compounds over time. The data is yours — exportable, portable, queryable.
Connect your data
Calendar, email, notes, habits. Your agent sees your actual life — not a blank slate.
Watch it learn
Every interaction enriches your personal dataset. Context compounds. Patterns emerge. Your agent gets smarter.
You own the output
Export it, query it, take it anywhere. Built on an open standard. No lock-in by design.
Not a settings page. Quests.
Your first quest: connect your calendar. Watch your agent prepare for tomorrow's meeting. Each step shows real value — not configuration.
Structured. Portable. Compounding.
A personal dataset that knows your preferences, patterns, decisions, and context. Queryable. Growable. Every session makes it richer. Every dataset makes your agent smarter.
Exportable in the Open People standard. Portable to any compatible system. Switch platforms anytime — your data comes with you. Not our training data. Yours.
"Every interaction makes your dataset richer. Every dataset makes your agent smarter. The compound effect is yours to keep — not ours."
Be first.
marsbot is in private beta. Join the waitlist and we'll reach out when your access is ready.
You're on the list. We'll be in touch.
Something went wrong. Try again.
No spam. We'll contact you personally.
Building in the open.
Three phases from foundation to launch. Progress is real — not vaporware.
Agent Foundation
Core agent runtime, tool system, context capture, and multi-surface deployment. The infrastructure is built.
Gamified Onboarding
First-time experience, quest system, data connectors. Making the agent useful from minute one.
Structured Dataset + Export
Personal identity graph, memory persistence, portable export. Your data becomes truly yours.
Building in the open
Every milestone, feature, and integration — documented as it ships.
TUI Daily Driver & Full Pipeline Verified
Terminal UI reaches daily-driver quality with 7 full views, integrated chat, keybindings, command palette, and gateway wiring. End-to-end pipeline verified at ~17s latency.
- TUI: 7 views (dashboard, threads, chat, tools, skills, agents, logs), full chat with streaming, keybinding system, command palette, gateway connection
- Gateway end-to-end pipeline verified at ~17s average latency across full agentic loop
- All P0 items complete — core runtime, chat, tools, providers, CLI, TUI
- All P1 items complete — bundled skills, sandbox, rate limiter, approval flow, cost tracking
- Desktop Electron app deferred — TUI covers the use case; Electron to return post-v1
Encrypted Vault & Security
Full encrypted credential vault with 7 tables, 31 passing tests, and a standalone security toolkit. Desktop settings surface vault management UI.
- Vault: 7 tables (credentials, keys, audit_log, shares, policies, snapshots, recovery), 31 tests, AES-256-GCM encryption at rest
- Desktop Settings: vault card UI with create/read/rotate/delete credentials and audit log viewer
- Security toolkit: 11 modules including prompt shield, SSRF guard, exec safety, and security audit runner
- Standalone redblue CLI for security audits, credential scanning, and policy enforcement outside the main gateway
P2 Quality Sweep
Persistence, spatial canvas polish, offline mutations, cost tracking, and plugin loading — the full P2 quality pass bringing production-readiness to core surfaces.
- Layout persistence: localStorage for instant restore, Supabase for cross-device sync, debounced writes
- Spatial canvas: drag-and-drop panels with snapping, resize handles, z-index management, panel registry
- Offline mode: mutation queue with retry, conflict resolution, sync status indicator
- Cost tracking: per-model token counting, session spend, budget cap enforcement with alerts
- Plugin loading: dynamic plugin resolution in gateway, hot-reload support, plugin sandboxing
Core Experience Wiring
Skills, sandbox, rate limiting, agent display names, and approval flow all wired up and test-covered. The core agent loop is now production-quality.
- Skills: 22 bundled skills, prompt builder with slot injection, skill registry with search, 16 integration tests
- Sandbox: 5 routable tool categories (shell, file, web, code, data), isolation per execution, 16 tests
- Rate limiter: token bucket algorithm, per-user and per-model limits, burst allowance, 24 tests
- Chat responses: agent display names surfaced in all surfaces (web, TUI, mobile), thread metadata enrichment
- Approval flow: 14 tests covering tool approval request, user decision, timeout, and escalation paths
P1 Tools & Bundled Skills
New high-value tools, bundled skill packs, and a hardened security layer with prompt shield and SSRF guard.
- New tools: web_search, apply_patch, edit_file, read_file, list_dir, run_shell, fetch_url, summarize, extract_json
- Bundled skills: coding assistant, research analyst, security auditor, data extractor, writing helper
- Prompt shield: injection detection, jailbreak pattern matching, policy enforcement before tool dispatch
- SSRF guard, exec safety wrappers, and security audit runner integrated into the gateway tool pipeline
Ship v1 Roadmap & Full Audit
Full audit of the platform, published ship-v1 roadmap, documentation reorganization, and public landing page. The project is now open and on a clear path to v1.
- Audited: 95 tools across all packages, 6 LLM providers, 6 channel adapters — inventory complete
- Published ship-v1 roadmap with P0/P1/P2 tiers, owners, and deadlines
- Reorganized docs: consistent structure across all 22+ packages, removed stale content
- Archived 25+ internal design docs to history/, keeping active docs lean
- Public landing page live at marsbot.dev — changelog, docs, and project overview
TUI Client & Gateway Session Manager
Terminal UI with command palette, tool catalogs, and event streaming. Gateway gains full session lifecycle management and the CLI can now start, stop, and inspect the daemon.
- TUI: command palette (Ctrl+P), tool catalogs with search, live event streaming panel, thread list with preview
- Gateway session lifecycle: create, resume, expire, list — sessions persist across reconnects
- Gateway config: hot-reloadable config file, per-session overrides, environment variable resolution
- CLI commands: gateway start, stop, status, logs, reload — daemon management from the terminal
Platform Packages Complete
All core platform packages shipped: gateway daemon, channel system, skills loader, CLI, Docker sandbox, plugin system, and workflow engine.
- Gateway daemon: JSON-RPC 2.0 over HTTP and WebSocket, request queuing, worker assignment, health endpoint
- Channel system: 6 adapters (web, mobile, CLI, TUI, Slack, webhook) with unified message envelope
- Skills system: SKILL.md loader with slot definitions, parameter validation, and prompt compilation
- CLI binary: 8 commands (chat, run, agent, tool, skill, gateway, config, version) with help and autocomplete
- Docker sandbox: per-execution containers, resource limits, volume isolation, output capture
- Plugin system: npm-loadable plugins, capability declaration, lifecycle hooks (init/teardown)
- Workflow engine: step graph, conditional branching, parallel fan-out, result aggregation
MCP Integration & Rate Limiting
Model Context Protocol bridge with 10 Supabase tools, multi-server MCP client management, per-server approval policies, and full rate limiting with health tracking.
- MCP bridge: 10 Supabase tools exposed via MCP protocol, bidirectional tool routing
- McpClientManager: spawn and manage multiple MCP servers, stdio and SSE transports
- Per-server approval policies: allow/deny/ask per tool category, configurable in gateway config
- Rate limiter: token bucket algorithm, per-user and per-server limits, queue with backpressure
- Health tracking: latency histograms per MCP server, error rates, automatic circuit breaker
Spatial UI & Agent Runtime v2
Spatial UI with draggable panels and Spotlight search. Agent runtime rebuilt from scratch with 13 tools, retry logic, approval gates, post-task reflection, and Monaco code editor.
- Spatial UI: draggable, resizable panels with snap-to-grid, Spotlight overlay (Cmd+K) for global search and commands
- Agent runtime v2: 13 built-in tools, configurable retry with backoff, approval gates per tool category
- Post-task reflection: agent writes a short memory entry after each completed task for future context
- Hybrid classifier: routes messages to tools vs. plain chat using intent scoring + keyword heuristics
- Code editor: Monaco embedded in thread view with syntax highlighting, diff view, and apply-to-file action
Desktop App & Provider Abstraction
Electron desktop app with IPC security and tray integration. Provider abstraction layer supporting 6 LLM backends. Full 7-page desktop UI.
- Electron app: secure IPC with context isolation, AES-256 credential storage, system tray with quick actions
- 6 LLM providers: Anthropic, OpenAI, Gemini, Groq, Ollama (local), OpenRouter — unified provider interface
- Desktop: 7 pages (chat, agents, tools, skills, settings, vault, logs) with persistent sidebar navigation
- Chrome extension: side panel mode, context injection from active tab, selection-to-prompt shortcut
Foundation
The core platform is alive: agent runtime v1 with agentic loop, 4 built-in tools, Supabase schema, React Native mobile app, and Realtime worker process.
- Agent runtime v1: agentic loop with tool dispatch, message history, streaming responses
- 4 built-in tools: web_fetch, run_code, read_memory, write_memory — the essential starting set
- Supabase schema: threads, messages, agents, tools, memories, events, audit_log tables with RLS
- Mobile app: React Native with Expo, thread list, chat view, agent selector, and settings
- Worker process: Supabase Realtime subscription, request claiming with optimistic lock, response delivery
"Corporations built empires on structured data about you.
We're giving you the same thing — and making it portable enough to leave the planet."
Phase 1 is your personal agent. The standard underneath was designed to travel.