Team-X
Run an AI company. Not a prompt.
Open-source, privacy-first, local-first desktop app for running AI-agent organizations. You don't manage prompts or pipelines; you run a company: hire employees from a curated role library, build an org chart with real hierarchy, set goals, break them into projects, file tickets, schedule future work, watch the team work in real-time, chat with anyone on demand, and pull everyone into an all-hands meeting with one click.
Download | Quick Start | User Guide | What's New | Contributing | Changelog
https://github.com/user-attachments/assets/bda8d55f-831b-465d-b361-ffe18bc54418
A real session on the shipped v3.5.0 build in Night Ops: hire a VP of Engineering, check the org chart, open Tickets from the Ctrl+K command palette, file tickets with owners, and watch Mission Control. The whole app is built on the Command Console design system: brushed-aluminum faceplates, phosphor LCD readouts, stencil word-lamps, and data-bound VU meters.
What's new
v3.5.0 is about trust: every setting now does what its label says.
- Enforced privacy tiers. Every model call, embedding and external runtime is checked against Settings → Privacy. A provider above the tier is refused with a reason and is never silently swapped.
- Copilot grounded in your company's knowledge, with long-term memory and a knowledge graph that survive restarts.
- The Models tab. Run GGUF models locally, browse Hugging Face, and connect LAN model servers.
- A palette that understands commands, plus meeting minutes that file action items, a Hire dialog with all 55 roles, and a per-company Proactive Mode that works.
- A Linux AppImage with no FUSE 2 requirement.
Full release notes: What's New · installers: v3.5.0 release.
Features
The Org
- 57 curated F10 roles: 55 user roles across 6 hierarchy levels (Officer, Senior Management, Management, Supervisor, Lead, IC) plus 2 hidden system roles, hand-written role specifications, not generic templates
- Multi-company workspace: run multiple AI organizations side-by-side, each with its own employees, goals, and settings
- Org chart editor: drag-to-rearrange reporting lines, promote, fire, set managers, edit employee profiles, visualize the full hierarchy with color-coded levels
- Hire dialog: searchable role catalog with level filter chips, one-click hiring from the curated pack
The Live Cockpit
- Mission Control dashboard: five subviews: Mission Control (live ops stats, queues, runs, runtime heartbeats), Timeline (event feed), Stream (raw LLM output), Floor (employee activity grid), Commands (palette history with intent labels)
- Annunciator rail: five real signal lamps (QUE / GGUF / BUDG / APPR / MTG) with a master-caution acknowledge ritual: blinking means unacknowledged, click to ack, click again to teleport to the source view
- Goals and projects: set company-level goals with auto-derived progress meters, create projects with lead / priority / target date, track everything on a status kanban
- Schedule and calendar: a team week view plus agenda: manual tasks and reminders alongside automatic ticket due dates, project targets, goal targets, overdue counters, and assigned agent wakeups
- Ticket kanban: 4-column board (Open, In Progress, Blocked, Done) with drag-to-move, a detail rail for participants / attachments / discussion / ticket memory, and automatic agent assignment
- One-click meetings: call an all-hands with selected attendees, interject mid-meeting, auto-generated minutes with action item extraction
- Real-time telemetry: company stats, daily usage charts, per-employee breakdown, cost analysis by provider and model with date range filtering plus Work / Agentic / Copilot run-kind filters
The Command Console design system
Team-X's interface is a purpose-built hardware design language, Command Console / Carbon Pro, modeled on professional broadcast and DJ gear rather than generic SaaS chrome:
- Dual-shift theming: Night Ops (brushed black aluminum, default) and Day Shift (silver anodized), switchable from the top bar. Displays stay void-black in both shifts, exactly like real silver hardware keeps black LCDs
- Stencil word-lamps: status is a word in a lamp tile (
GO / HOLD / NO-GO / STBY / EXEC), never an ambiguous icon or colored dot
- Functional instrumentation: LCD readouts and VU meters are data-bound only (queue pressure, goal progress, workforce utilization); nothing blinks or glows for decoration
- The dual-form red rule: steady red means LIVE / armed / command authority; blinking red is a question that demands an answer
The same dashboard in Day Shift: silver chassis, displays stay dark.
Autonomy control plane
A full operator console for governed autonomous work (the Autonomy tab, 10 subviews):
- Doctor: pre-flight health workflow: DB integrity, recovery readiness, runtime posture, secrets, provider/MCP health, budget blockers
- Benchmarks: replay deterministic autonomy scenarios and inspect pass rates, recovery timing, duplicate-work prevention, and spend
- Runtime profiles & operations: bind execution profiles (internal / local / external agents) to employees; watch live heartbeats and ticket checkouts
- Routines: recurring operating loops that materialize into explicit, auditable ticket runs
- Budgets: spend governance with warnings, hard caps, escalation thresholds, and a full ledger across company / employee / runtime / routine scopes
- Approvals: one unified decision inbox for authority, planner, budget, and routine decisions
- Artifacts, Memory, Access, Improve: runtime evidence, thread digests + resumable checkpoints, operator/sharing posture, and a self-improvement loop that opens deduped correction tickets from recent failure signals
AI Runtime
- 9 LLM providers: Ollama (local), Anthropic, OpenAI, Google Gemini, Groq, OpenRouter, Together AI, Fireworks AI, plus any OpenAI-compatible endpoint
- Privacy tiers: Local, Open-Source Cloud, Proprietary Cloud; filter which providers your agents can use
- Adaptive runtime strategy: Auto, Hybrid, Always-On, or Lean mode based on a live hardware profile (CPU cores, RAM, GPU, VRAM shown in Settings). GPU detection is native on all three platforms:
wmic on Windows, system_profiler on macOS (Apple Silicon reports its unified-memory GPU budget), and nvidia-smi with an lspci fallback on Linux
- MCP tool calling: agents use Model Context Protocol tools via a singleton host with connection pooling and
tools_allowed/tools_denied enforcement
- Skills & authority: install skills from a local folder or GitHub URL, import MCP servers from templates or manual config, and govern everything through authority grants and per-employee permissions
- Employee-to-employee messaging: agents communicate with colleagues via built-in tools, forming collaborative workflows
- Local GGUF models: a Models tab with four panels: Library (add a
.gguf file or watch a folder, rescan or stop watching it, per-model system prompt / chat template / advanced tuning, load and unload), Discover (search Hugging Face, browse a repo's files, queue resumable downloads that survive a pause or a quit), Endpoints (LM Studio / Ollama / llama-server / KoboldCPP / vLLM on your LAN, validated as local-network only and probed for reachability), and Runtime (GPU inventory across CUDA / ROCm / Vulkan / Metal / CPU, active backend, LRU pool capacity, bundled llama.cpp build). Per-model benchmarks record prompt-eval and generation throughput measured from llama-server's own timings, never an estimate. A GGUF model is not yet an agent provider; Ollama remains the local path for employees
Intelligence Layer
- RAG-grounded agent turns: every agent prompt is augmented with retrieved context from messages and vault files via the
@team-x/intelligence package (embeddings stored as BLOBs in SQLite and ranked in-process by cosine similarity: an exact per-company scan, or an IVF approximate index once a company holds 4,096 or more vectors; token-aware chunking with overlap, cosine-threshold gating, SHA256-dedup attribution blocks)
- Enhanced AI subsystem: semantic chunking for the RAG index (splits on headings, paragraphs and code fences instead of a blind character window), query expansion, long-term memory with LLM fact extraction, a cross-thread knowledge graph (both persisted in SQLite), and sampled span tracing. It is active whenever RAG is enabled with an embedding provider, grounds Copilot answers through the
search_company_knowledge tool, and remembers completed Copilot exchanges. Each is a toggle in Settings → Enhanced AI; the Copilot's own model writes the answer from that grounding and cites its sources
- Natural-language command palette (
Cmd+K): 14 structured intents (hire / fire / promote / assign / create / close / reopen / project / goal / meeting / status / navigation / vault search) plus a complex_request fallback that hands off to the agentic loop. LLM-backed classifier with JSON-output retry, fuzzy entity resolution, FTS5 ticket lookup, destructive-action confirmation gate, last-20 command history
- Agentic loop for complex questions: ask free-form questions like "why is the frontend team behind schedule?" and get a grounded multi-paragraph answer citing specific tickets, employees, and events. Runs a ReAct-style loop on a hidden
system-agent pseudo-employee, dispatches six read-only query tools, and terminates under hard step / token / wall-clock budgets (defaults 8 / 8000 / 120s, configurable in Settings → Agentic Loop)
- Live step log + persisted thread: the palette streams each loop step as a labeled card (plan → tool call → tool result → answer) with provider and token footer; full transcripts persist for later reference, and every run can be canceled mid-flight
- Task Planner (write-side): Management-and-above agents can decompose projects into tickets, delegate subtasks with deterministic workload scoring, and review deliverables. Level-gated tool injection, an amber confirmation gate before any ticket is created, and four clamped guardrail settings in Settings → Task Planner
- Copilot (proactive analyst): a hidden
system-copilot employee analyzes the company on a 5-minute cadence (plus event triggers) and surfaces insights across five categories (operational / cost / org / / ) with three severities. Deterministic dedup, dismissal-feedback weight tuning (advisory by default, self-tuning opt-in), and zero phone-home: analysis runs on whatever provider you've configured
Ship-Ready
- File vault + agent-created deliverables: filesystem-backed blob storage with SHA256 integrity verification, FTS5 full-text search, ticket attachments, and employee-generated
txt / md / csv / json / html / docx / xlsx / pptx outputs
- One-click backup/restore: full SQLite + vault archive with manifest validation
- Append-only audit log: filterable event timeline with summary cards, actor search, date range picker, and CSV/JSON export
- Workspace portability: export a company as a template package, preview and import packages, share with a reviewed posture
- Paperclip Import: preview a Paperclip export folder (agents, runtimes, tickets, skills, unsupported adapters, missing secrets), save it as a
.teamx-package.json, and review & import it through Portability; nothing is created until Portability commits it
- Private Operator Access: a read-only decision record of what a non-workstation device would be allowed to do in a workspace (Localhost / Tailscale / Hosted bridge), with every refusal and its reason; nothing opens a listener
- In-app User Guide: role-based onboarding (owner / operator / builder) with 25 sections, progress tracking, and deep links that jump straight into the referenced view or settings section
- Cross-platform installers: Windows (NSIS), macOS (DMG), Linux (AppImage + .deb) via electron-builder
- User-triggered updates: check for new versions from GitHub Releases on demand (zero phone-home)
Installation
Download
Grab the latest release for your platform from GitHub Releases:
| Platform | File | Architecture |
|---|
| Windows | Team-X-x.x.x-Setup-x64.exe / Team-X-x.x.x-Setup-arm64.exe | x64, arm64 |
| macOS | Team-X-x.x.x-x64.dmg / Team-X-x.x.x-arm64.dmg | x64 (Intel), arm64 (Apple Silicon) |
| Linux | Team-X-x.x.x-x86_64.AppImage / Team-X-x.x.x-amd64.deb | x64 |
Linux AppImage note: v3.5.0 and later embed AppImage's static runtime, so the AppImage starts on a stock Ubuntu 22.04 / 24.04 desktop with no extra packages (#16). v3.4.0 and earlier need FUSE 2 (libfuse2; libfuse2t64 on Ubuntu 24.04; fuse-libs on Fedora); if one of those won't start, install that package or run ./Team-X-x.x.x-x86_64.AppImage --appimage-extract-and-run. The .deb never needs FUSE and resolves its own dependencies (sudo apt install ./Team-X-x.x.x-amd64.deb).
Unsigned builds: the installers are not code-signed yet, so Windows and macOS warn before the first run. Check a download against the release's SHA256SUMS.txt; from v3.5.0 you can also verify its build provenance with gh attestation verify <file> --repo Git-Rocky-Stack/Team-X.
From Source
# Prerequisites: Node.js 22.13+ (repo pins 22.22.2 via .nvmrc), pnpm 9+
git clone https://github.com/Git-Rocky-Stack/Team-X.git
cd Team-X
pnpm install
pnpm dev
Local LLM Setup (Ollama)
Team-X works fully offline with local models. Install Ollama, then:
ollama serve
ollama pull llama3.1:8b
Launch Team-X. Its first launch adds an enabled Ollama (Local) provider at http://localhost:11434, so a running Ollama needs no further setup.
Cloud Providers
Add any supported provider in Settings > AI Providers: enter your API key, test the connection, and toggle it on. Your keys are stored in the OS keychain (never in config files).
Architecture
Team-X/
apps/desktop/ Electron app
src/main/ Main process (Node.js + TypeScript)
db/ SQLite + Drizzle ORM (38 migrations)
ipc/ Typed IPC handlers (236 registrations / 229 bridge methods)
orchestrator/ Agent scheduler + event bus
services/ Vault, backup, MCP host, providers, updater,
rag-indexer, command-service, agentic-loop,
copilot-analyzer, routine-service, budgets,
local-gguf (library / pool / runtime / endpoints /
hf-downloads / benchmarks), heartbeat
src/preload/ Context-isolated bridge (TeamXApi)
src/renderer/ React 19 + Tailwind + shadcn/ui on the
Command Console design system (DESIGN.md)
components/console/ Faceplate, LampTile, LcdWell, MetricTile,
VuMeter, AnnunciatorRail, ShiftToggle, Tag
features/ audit, autonomy, chat, command, copilot,
dashboard, hire, meetings, memory, models,
orgchart, projects, settings, telemetry, tickets,
user-guide, vault, workspace
hooks/ React Query hooks
store/ Zustand app store
e2e/ 20 Playwright specs / 29 cases
packages/
shared-types/ IPC contract types, event types, entities
role-schema/ Role-spec parser + renderer + pack signing
provider-router/ LLM provider registry + streaming adapters
telemetry-core/ Cost calculation utilities
intelligence/ RAG (chunker, embedding, retriever),
NLU (classifier, entity resolver, slot filler),
agentic loop (ReAct scheduler, tool registry)
local-gguf-runtime/ GPU probe, llama.cpp lifecycle, LRU pool,
GGUF parser, folder watcher, auto-tune
role-packs/
strategia-official/
roles/ 57 curated F10 roles across user + system levels
roles/system/ Hidden system-agent + system-copilot role cards
Key Design Decisions
- Renderer is a pure view. No direct LLM or MCP calls. Every interaction crosses the typed IPC bridge.
- Orchestrator is the only scheduler. Pause semantics (e.g., during meetings) are race-free because nothing dispatches without the orchestrator's consent.
- MCP Host is a singleton. One pool of connections shared across all agents, with no N-client sprawl.
- Storage is SQLite + filesystem. Metadata in SQLite, blobs on disk, SHA256 integrity. File blobs never go in the database.
- Provider resolution is the single LLM gateway. The provider factory and runtime-profile resolver enforce the privacy tier on every run; the provider router carries concurrency caps and cost tracking.
- Events table is append-only. Source of truth for the real-time dashboard and audit log.
- The design system is law. Every surface follows
DESIGN.md (Command Console / Carbon Pro): four-layer hardware depth, dual-shift theming, displays-stay-dark, data-bound instrumentation only.
- Zero phone-home. Ever. No telemetry, no analytics, no auto-update checks. Updates are explicitly user-triggered.
- Secrets live in the OS keychain via keytar. Never in config files.
Tech Stack
| Layer | Technology |
|---|
| Desktop shell | Electron 44 |
| Build | electron-vite + electron-builder |
| Main process | Node.js 22 + TypeScript (strict) |
| Renderer | React 19 + TypeScript + Tailwind CSS + shadcn/ui + Radix |
| Design system | Command Console / Carbon Pro (DESIGN.md): Archivo, Public Sans, Departure Mono, Iosevka |
| State management | Zustand + React Query |
| LLM integration | Vercel AI SDK + provider packages |
| Agent framework | Custom orchestrator (not LangChain/CrewAI) |
| MCP | @modelcontextprotocol/sdk |
| Local models | Ollama as the local agent provider; native GGUF via @team-x/local-gguf-runtime (llama.cpp b9371) and the Models tab |
| Database | better-sqlite3 + Drizzle ORM (38 migrations) |
| Full-text search | SQLite FTS5 |
| Secrets | keytar (OS keychain) |
| Package manager | pnpm workspaces |
| Lint / format | Biome + ESLint |
| Unit tests | Vitest (4,718 tests / 376 files) |
| E2E tests | Playwright (20 specs / 29 cases) |
| CI | GitHub Actions (Ubuntu + macOS + Windows + Electron E2E smoke) |
Documentation
Team-X includes comprehensive documentation across 142 files:
Getting Started
- Quick Start Guide: 15-minute setup walkthrough for new users
- Getting Started: first company, first hire, first ticket
- Comprehensive User Guide: complete product documentation
- In-app User Guide: role-based onboarding with progress tracking, built into the app (left rail → User Guide)
Reference
Real-World Scenarios
Templates
- Templates: handoff documents, meeting agendas, ticket templates, routine templates
Developer
- Developer Reference: architecture, MCP server development, role pack format, internal IPC surface
- Integration Guide: providers, MCP servers, and extension points
- Dependency Upgrade Program: the audit gate, accepted exceptions, and the order of the remaining major upgrades
- Release Signing: signing credentials, the signature gate, install smoke tests, provenance, and dry runs
- Branch Governance: the
main and release-tag rulesets, the review wall, break-glass, and signed tags
- Architecture · API Endpoints · Database Schema · Design System
Advanced
AI Discovery
- llms.txt: AI-optimized documentation index for LLM consumption
- long-llms.txt: extended AI reference with comprehensive details
Development
# Install dependencies (runs electron-rebuild postinstall)
pnpm install
# Start dev server with HMR
pnpm dev
# Run unit tests
pnpm test
# Run E2E tests (builds first)
pnpm -F @team-x/desktop test:e2e
# Typecheck all workspaces
pnpm typecheck
# Lint + format
pnpm lint
pnpm format
Building Installers
pnpm dist # Current platform
pnpm dist:win # Windows NSIS (x64 + arm64)
pnpm dist:mac # macOS DMG (x64 + arm64)
pnpm dist:linux # Linux AppImage + .deb (x64)
Database Migrations
pnpm -F @team-x/desktop exec drizzle-kit generate --name <snake_case_name>
See CONTRIBUTING.md for the full development guide.
Testing
Team-X ships with 4,718 unit tests across 376 files and 20 Playwright E2E specs (29 cases):
| Spec | Coverage |
|---|
smoke.spec.ts | Boot, render employees, chat round-trip |
ticket-flow.spec.ts | Create ticket, assign, agent reply |
meeting-flow.spec.ts | Call meeting, interject, end, verify minutes |
vault-backup.spec.ts | Vault upload, integrity check, backup create/verify |
rag-flow.spec.ts | RAG retrieval and attribution in agent turns |
command-palette.spec.ts | Cmd+K intent classification, destructive gate, history |
agentic-loop.spec.ts | Complex-request agentic loop, step log, persisted thread |
task-planner.spec.ts | Write-side planner, amber confirmation gate, decompose → delegate round-trip |
copilot-service.spec.ts | Periodic analyzer tick, insight dedup, dismiss, ask-the-copilot, gate regression guards |
copilot-ui.spec.ts | Sparkles toolbar toggle, sidebar + insight card, dismiss optimistic update, Cmd+Shift+K shortcut |
copilot-feedback.spec.ts | Dismissal feedback, advisory weight suggestion, audit event, Settings weight display |
copilot-insight-export.spec.ts | Sidebar category filter, company-scope JSON export, all-company CSV export |
phase-5-integration.spec.ts | Cross-milestone stitch: RAG indexing, palette round-trip, write-side gate, copilot tick |
phase-6-integration.spec.ts | Planner role-fit outcome, feedback apply, run-kind telemetry filters, local JSON export |
workspace-switcher.spec.ts | Multi-company switcher, create/settings/archive/delete lifecycle, hire flow |
org-chart.spec.ts | Org chart read-side, drag reporting-line reassignment, cycle rejection, promote/fire |
telemetry-kind-filter.spec.ts | Per-kind telemetry totals for All, Copilot, and Work run buckets |
local-gguf-loading.spec.ts | Local GGUF runtime settings, GPU inventory, pool status round-trips |
All E2E specs run against a canned test-mode provider: no Ollama, no API keys, no network.
pnpm test # Unit tests
pnpm -F @team-x/desktop test:e2e # E2E (full build + Playwright)
Privacy
Team-X is built with a privacy-first posture:
- Local-first by default. Runs entirely on local models (Ollama) with no cloud dependency.
- Zero phone-home. No analytics, no telemetry, no crash reporting, no auto-update checks.
- OS keychain for secrets. API keys are stored in the system keychain via keytar, never in plaintext config files.
- Enforced privacy tiers. Choose which provider tiers (local, open-source cloud, proprietary cloud) may be used. Every model call and external runtime is checked; a provider above the tier is refused with a reason, never silently swapped.
- Your data stays yours. All data lives in a local SQLite database and filesystem vault on your machine.
Contributing
We welcome contributions. See CONTRIBUTING.md for development setup, PR guidelines, coding standards, and the role-pack contribution guide.
License
MIT © 2026 Rocky Elsalaymeh
