tangerine-teams-app

# Tangerine Teams App > **The team work memory layer no one else does.** > > Tangerine captures **two layers** that no other tool combines: > 1. **AI tool work** — Cursor / Claude / ChatGPT / Codex / Windsurf / Gemini / Devin / Replit / Apple Intelligence / MS Copilot > 2. **Human communication** — Discord (real) · Slack / GitHub / Linear / Calendar (beta) · Lark / Zoom / Teams (coming v1.13+) > > Every other tool only does **one** layer (or neither, well): > > | Tool | AI tool work | Human comm | Position | > |---|---|---|---| > | **Tangerine** | ✅ 8 vendors | ✅ Discord real + 4 beta | **Unique cross-section** | > | Obsidian | ❌ | ❌ | Personal markdown | > | Notion | ❌ | partial | Personal/team docs | > | Mem.ai | partial | ❌ | Personal AI notes | > | Glean | ❌ | partial | Closed enterprise search | > | Linear / Slack / GitHub | ❌ | self-only | Single-channel | [](https://github.com/Tangerine-Teams-Intelligence/tangerine-teams-app/releases/latest) [](PRIOR-LICENSE-APACHE-2.0.txt) [](LICENSE) **Status**: v1.12.1 — dual-layer memory · auto-updater shipped · Lark integration coming v1.13 --- ## What we sell that no one else has **Your team's combined work-trace** — what your team's AI tools did + what your team said to each other → unified, searchable, AI-readable, in your own git repo. Without Tangerine: 5 silos (Cursor history / Claude session / Slack thread / GitHub PR / your notes). Your AI tools see only their own slice. Your team forgets what was decided 3 days ago. With Tangerine: 1 markdown brain + 1 git repo. Every AI tool can read team context. Every team member sees what their AIs did + what other humans said. Search "what did we decide on PCB" returns relevant atoms across all sources. <!-- === wave 25 === Hero screenshots — placeholders point to dogfood-prep/video/screenshots/ --> <!-- Replace with absolute GitHub raw URLs once images are committed to the repo. --> |  |  |  | |---|---|---| | Dashboard with shared memory | Memory tree + atom preview | Memory graph view | <!-- === end wave 25 === --> --- ## Why Tangerine — 4 design moats, each with concrete user value ### 1. Use your existing Cursor / Claude / ChatGPT **No new $20/seat/month for our LLM.** You already pay Cursor Pro / Claude Pro / Codex / ChatGPT Plus. Tangerine reverse-calls into your editor through the **MCP sampling protocol** — your existing LLM does the inference, we just route the prompt. Your team pays $0 in extra LLM fees. If your team has no AI subscription, our DeepSeek-backed fallback runs ~$2-5/team/month flat (not per-seat, not per-token). How it works (Tauri ⇄ MCP server ⇄ your editor): ``` ~/.cursor/mcp.json Tangerine WS server │ 127.0.0.1:7780/sampler ▼ ▲ MCP server ────── register_sampler ────────────────────── │ │ │ ◄────── sample (prompt + system) ───────── Tauri co-thinker │ server.createMessage() (reverse-call to Cursor's host LLM) │ └────── sample_response (LLM output) ──────────────► back ``` ### 2. AGI brain is a markdown doc — readable, editable, git-able **No black box. You see exactly what the AGI thinks.** Open `~/.tangerine-memory/team/co-thinker.md` any time. It's a markdown doc — `cat` it, `vim` it, `git diff` it, push it to your repo. The AGI's "thinking" is a document you can read. - Heartbeat: every 5 min when team's idle, every 30 min when active - Don't like what the AGI wrote? Edit the markdown directly. It picks up your edits next heartbeat. - Diff your AGI's brain across weeks: `git log -p co-thinker.md` ### 3. Cross-vendor AI tool visibility **Anyone on your team uses any AI tool — we capture it.** Your designer uses Cursor. Your CTO uses Claude Code. Your contractor uses ChatGPT. Your intern uses Replit. Most tools today see only their own slice. Tangerine reads each vendor's local conversation files and unifies them into one team memory. 8 vendor parsers shipped: | Vendor | Status | Local files read | |---|---|---| | Claude Code | Confirmed working | `~/.claude/projects/<repo>/*.jsonl` (41 sessions / 8483 messages tested) | | Cursor | Wired, awaiting real-files validation | `~/.cursor/conversations/*.json` | | Codex | Wired, awaiting validation | `~/.codex/sessions/*.json` | | Windsurf | Wired, awaiting validation | `~/.windsurf/conversations/*.json` | | Devin | API-based, needs real token | Cognition API | | Replit | API-based, needs real token | Replit API | | Apple Intelligence | macOS-only, awaiting validation | `~/Library/Application Support/com.apple.intelligenceplatform/` | | MS Copilot | Wired, awaiting validation | M365 Copilot API | Honest disclosure: only Claude Code is real-files-validated as of v1.9.1. The other 7 parsers compile, have unit tests, and have spec-correct schemas — they just haven't been pointed at real user files yet because we (the dev team) don't have those tools installed. As users install Tangerine + their respective tools, we close each row. ### 4. AI tools as first-class sidebar **Most apps' sidebars hold workflows. Ours holds AI tools — because that's where the work actually happens.** Cursor / Claude Code / Codex / Windsurf / ChatGPT / Devin / Replit / Apple Intelligence / MS Copilot / Ollama — all in one sidebar. - Star your primary tool (it's where new tasks default to) - See each tool's active session indicator at a glance - One-click "Copy MCP config to clipboard" for tools that support MCP - Auto-config card surfaces for installed MCP-compatible tools --- <!-- === wave 25 === Why Tangerine vs Obsidian comparison --> ## Why Tangerine vs Obsidian Obsidian is the gold standard for personal markdown knowledge management. Tangerine is "Obsidian for teams + AI tools." We borrow the best of Obsidian (markdown files you own, graph view, plugin ecosystem) and add the team + AI surface Obsidian deliberately doesn't ship. | Capability | Obsidian | Tangerine | |---|---|---| | Markdown files you own | Yes | Yes (`~/.tangerine-memory/`) | | Graph view of notes | Yes | Yes (`/memory` graph toggle) | | Daily notes | Plugin (Calendar) | Built-in (`/today`) | | Per-user vs shared scope | Single vault | `team/` + `personal/<user>/` split | | AI tool sidebar | None | 10 tools, MCP auto-config | | Cross-vendor AI memory | None | 8 parsers (Cursor / Claude Code / Codex / Windsurf / Devin / Replit / Apple Intelligence / MS Copilot) | | AGI co-thinker brain | None | `team/co-thinker.md` heartbeat-driven | | MCP sampling protocol | None | Reverse-call your editor's LLM | | Team sync | Plugin (git, paid Sync) | Built-in (git mirror, Cloud) | | Auto-updater (in-app) | Yes (manual download) | Yes (one-click install) | | Open source | Closed core, open plugins | AGPL v3 + Commercial dual | | Pricing for teams | $5/user/month (Sync) | $80/team flat (no per-seat) | | Linear / Obsidian style | Obsidian-first | "Linear UX, Obsidian data" | If your team already lives in Obsidian, you keep your Obsidian setup — Tangerine reads any markdown vault you point it at. We're additive, not replacement. <!-- === end wave 25 === --> --- ## Quick start (30 seconds from install) 1. Download installer for your OS from the [latest release](https://github.com/Tangerine-Teams-Intelligence/tangerine-teams-app/releases/latest): <!-- === wave 17 === cross-platform installer matrix shipped v1.10.4 --> | Platform | Architecture | Asset | Notes | |---|---|---|---| | Windows 10/11 | x64 | `Tangerine.AI.Teams_<v>_x64-setup.exe` | Full feature set incl. meeting capture | | macOS 11+ | Apple Silicon | `Tangerine.AI.Teams_<v>_aarch64.dmg` | Unsigned beta — **right-click → Open** the first time to bypass Gatekeeper | | macOS 11+ | Intel | `Tangerine.AI.Teams_<v>_x64.dmg` | Unsigned beta — **right-click → Open** the first time | | Ubuntu 22.04+ / Debian | x64 | `tangerine-ai-teams_<v>_amd64.deb` | Shell preview — meeting capture comes in a future release | | Any glibc Linux | x64 | `Tangerine.AI.Teams_<v>_amd64.AppImage` | Shell preview — `chmod +x` then run | Replace `<v>` with the current release version. Cross-platform parity status: - Windows: full app (Discord meeting capture + memory + AI sidebar) - macOS / Linux: shell preview (memory viewer + AI sidebar work; Discord meeting capture pending — bundled-Python pipeline is Windows-only this release) 2. Install + launch 3. WelcomeOverlay walks you through 4 cards in ~30 seconds 4. Click "Get started" → auto-detects installed AI tools and surfaces one-click MCP config snippets 5. Initialize co-thinker brain → first heartbeat in ~5-10 seconds --- ## Pricing — transparent because we build in public Tangerine is **OSS forever**. Run it on your laptop, connect your own DeepSeek API key, never pay us anything. Every feature is open-source, every commit is public. There are no enterprise-only forks hidden behind a paywall. If you want zero-config + managed everything, four tiers: | Tier | Price | What you get | |---|---|---| | **Self-host** | **$0 forever** | Full app, every feature. Bring your own DeepSeek API key (~$0.27/M tokens at cost) — or borrow your editor's LLM via MCP sampling. Your existing Cursor Pro / Claude Pro / Codex subscription becomes Tangerine's inference engine. | | **Tangerine Founder** | **$50 one-time** | Lifetime access to the Tangerine Cloud feature set for the founding team. One purchase covers the team forever. Doors close once we hit 200 founder seats. | | **Tangerine Cloud** | **$80 / team / month flat**<br>+ overage credits | Shared DeepSeek API key managed by us. ~100M tokens included monthly (very heavy team usage). Above that, top up with credits at $5 / 10M tokens. **Not per-seat**: invite 5 people or 50, the base price stays $80. | | **Commercial Honor Tier** | **$25 / user / year** | For commercial users who want to support the project but stay self-hosted. Honor system — we don't gate features. Funds go to OSS development; you get a public thank-you on the supporters page. | | **Marketplace** | 82% to sellers / 18% to us | Buy & sell community templates, sources, agents. Sellers set prices; we handle distribution + auth + billing. | | **Enterprise** | **$50–200k / year** | SOC 2 Type II, SSO, on-prem deployment, white-label, 24h SLA, dedicated account manager. **Required if you redistribute Tangerine inside your products** — AGPL forces it; buy a commercial license to keep your modifications private. | ### Why $80/team flat instead of $20/seat? Per-seat pricing trains you to "save" by not inviting your full team. Tangerine **only works when everyone is in it** — that's how cross-AI alignment happens. So we charge per-team, not per-seat. Either everyone uses it or no one does. Math comparison for a 5-person team: - ChatGPT Team: $25/seat × 5 = **$125/month** - Linear + Notion + Slack combined: ~$25/seat × 5 = **$125/month** - **Tangerine Cloud: $80/month flat** ← still cheaper than either, *and* it covers the cross-AI alignment piece those tools don't For a 30-person team, you save 80% vs ChatGPT Team. For 100-person, 95%. Self-host is still $0. ### How the shared DeepSeek key works Cloud teams share **one DeepSeek API key** that we manage: - ~100M tokens included monthly (covers ~30 people of normal co-thinker + atom edits + heartbeat) - Above that → credit packs ($5 per 10M tokens, no markup tricks; DeepSeek raw is ~$2.7/10M) - **You see the meter live** in Settings → Billing. No "surprise bill at end of month." - If you'd rather use your own DeepSeek key (or your own LLM), Self-host is $0 forever. ### Three promises 1. **No per-seat trap.** Grow your team — your base bill doesn't grow. 2. **No paywall pivot.** Self-host stays open-source forever. If we ever raise prices, you have a fork in your back pocket. 3. **Profit in public.** Our ARR + cost dashboard goes live with v2.0. You'll see real numbers per tier, including DeepSeek margin. ### Why charge at all if it's open-source? Two answers: - **Time** — Cloud customers don't want to set up DeepSeek API keys, host an MCP server, manage SSL, run backups, patch security holes. $80/month buys their engineer's time back. - **AGPL gate** — If you redistribute modified Tangerine in a closed-source product, AGPL requires you to open-source your modifications. Most enterprises can't. They buy a commercial license to keep their changes private. That funds the OSS for the rest of us. See [BUSINESS_MODEL_SPEC.md](./BUSINESS_MODEL_SPEC.md) for the full economic breakdown. --- <!-- === wave 1.13-E === --> ## Local-first by default Tangerine runs 100% on your machines by default. Your AI tool history, your memory, your tokens — all local. We can't read any of it. Cloud sync is opt-in only ($80/team/mo) — even then, E2E encrypted. Concretely: - **Memory dir** (`~/.tangerine-memory/`) — local plain-markdown. - **Source tokens** (Lark / Zoom / Teams / Slack / GitHub OAuth tokens) — encrypted in your **OS keychain** (Windows Credential Manager / macOS Keychain / Linux Secret Service). The Privacy panel shows presence-only, never the value. - **Whisper transcription** — bundled `faster-whisper` runs on your CPU. - **MCP sampling** — uses your editor's LLM, not ours. - **Discord bot** — Node.js subprocess on your machine. - **AI tool conversations** — read locally from each editor's history dir, never uploaded. What may leave your machine: - **Git push** to your specified remote (your data → your server). - **Telemetry** (anonymized, opt-out via toggle in Settings → Privacy). - **Auto-updater check** (version comparison only, no data). - **Cloud sync** — opt-in only, E2E encrypted. The Privacy panel (Settings → Privacy) renders this as a data-flow diagram, lists what's local vs what may egress, surfaces the per-source token presence, and ships a one-click "Verify local-execution" button that audits the last hour of network calls. <!-- === end wave 1.13-E === --> --- ## Architecture Tauri 2 desktop app (Rust backend + React frontend), file-based memory in your home dir, optional MCP server for AI tool integration. ``` ~/.tangerine-memory/ Tauri app MCP server (Node) ├── team/ ├── React UI ├── 7 MCP tools (sources/sinks) │ ├── co-thinker.md ◄──heartbeat───┤ Zustand store │ │ ├── decisions/ │ i18next (en/zh) │ │ └── timeline/ ├── Tauri commands │ └── personal/ │ 185 typed ├── Sampling bridge └── <user>/ ├── ws server │ /sampler ws ├── threads/ │ 127.0.0.1:7780 ◄──────┤ reverse-call to host LLM └── tools/ └── faster-whisper │ (local transcription) ``` - **Sources**: Discord (live), GitHub, Linear, Slack, Calendar, RSS / Podcast / YouTube / Article (hand-rolled parsers) - **Sinks**: MCP server (7 tools), browser extension (smart inject on ChatGPT/Claude.ai/Gemini), local file write - **8 personal agents**: parsers for Cursor / Claude Code / Codex / Windsurf / Devin / Replit / Apple Intelligence / MS Copilot --- ## License **Currently effective**: [Apache-2.0](PRIOR-LICENSE-APACHE-2.0.txt). Copyright 2026 Tangerine Intelligence Inc. **Transition draft (pending CEO + legal counsel ratification)**: [AGPL v3 + Dual Commercial](LICENSE). See [COMMERCIAL-LICENSE.md](COMMERCIAL-LICENSE.md) for commercial terms. The license flip will be announced when ratification completes. Until then, the codebase is governed by Apache-2.0. --- ## Status (v1.12.0) | What | State | |---|---| | Tests | 548+ Rust + 410+ frontend + 8 Playwright + 92 MCP vitest = 1058+ tests pass | | MCP sampling protocol | Real, end-to-end | | Welcome tour | 4-card 30-sec onboarding | | Setup wizard | Chat-driven (Wave 18) — type your stack in one line | | Auto-updater | In-app banner + one-click install (Wave 25) | | Memory views | File tree + graph + brain split editor (Waves 21-23) | | Daily dashboard | `/today` with seeded sample atoms (Wave 13 + 20) | | 中文 i18n | 287+ strings × 2 locales (en/zh 1:1) | | Cross-vendor parsers | 1/8 real-files-validated (Claude Code), 7/8 wired-and-tested | | Cross-platform builds | 5 installers per release (Windows / macOS aarch64 + x64 / Linux .deb + .AppImage) | | Public release | v1.12.0 | See [V1_9_ACCEPTANCE.md](./V1_9_ACCEPTANCE.md) for the v1.9 quality gates that this release inherits + extends. --- ## Docs - [SUGGESTION_ENGINE_SPEC.md](./SUGGESTION_ENGINE_SPEC.md) — 6 anti-Clippy disciplines + 4 visual tiers + 7 templates - [V2_0_SPEC.md](./V2_0_SPEC.md) — visualization-first (graphs, lineage, social) - [V2_5_SPEC.md](./V2_5_SPEC.md) — decision review + paywall - [V3_0_SPEC.md](./V3_0_SPEC.md) — personal agents + external world - [V3_5_SPEC.md](./V3_5_SPEC.md) — marketplace + enterprise - [BUSINESS_MODEL_SPEC.md](./BUSINESS_MODEL_SPEC.md) — anti-SaaS economics - [COMPETITIVE_ADVANTAGE.md](./COMPETITIVE_ADVANTAGE.md) — execution-leader positioning vs concept-creators - [DATA_MODEL_SPEC.md](./DATA_MODEL_SPEC.md) — atom schemas - [API_SURFACE_SPEC.md](./API_SURFACE_SPEC.md) — Tauri command catalog - [VISUAL_DESIGN_SPEC.md](./VISUAL_DESIGN_SPEC.md) — design tokens - [OBSERVABILITY_SPEC.md](./OBSERVABILITY_SPEC.md) — error/perf/i18n/SOC2 --- ## Owner Daizhe Zou — daizhe@berkeley.edu. Tangerine Intelligence Inc. (Delaware C-Corp). Best-effort response on issues.