tangerine-mcp

Align every AI tool on your team with your team's actual workflow.

License — transition in progress (2026-04)

Tangerine is moving from Apache 2.0 to AGPL v3 + Dual Commercial License. See LICENSE (draft AGPL v3 + dual) and COMMERCIAL-LICENSE.md (draft commercial terms). Until ratification announced, codebase is governed by Apache 2.0 — see PRIOR-LICENSE-APACHE-2.0.txt.

Your team uses Cursor, Claude, ChatGPT — but each AI sees a different slice of what your team's actually doing. This MCP server is the Sink that aligns them: it exposes Tangerine team memory to Claude Code, Cursor, Claude Desktop, and any other MCP-compatible AI client.

Once installed, your AI tool gains 7 tools it can call autonomously — no copy-paste, no manual injection:

| Tool | Calls when… | |---|---| | query_team_memory(query, limit?) | the user asks about prior decisions, meetings, anything that might be in team context | | get_today_brief() | session start — pulls today's brief so the AI loads context before the user asks | | get_my_pending(user) | the user asks "what's on my plate", "what do I owe", "what's overdue" | | get_for_person(name) | "what's been working on", "catch me up on ", before a 1:1 | | get_for_project(slug) | "status of ", "what's happening with " | | get_thread_state(topic) | "where did we land on ", "is still open" | | get_recent_decisions(days?) | "what did we decide recently", session start to load decision context |

Ask Claude "what did we decide about Whisper?" → it auto-calls query_team_memory("Whisper"). Ask "what's on my plate?" → it auto-calls get_my_pending("daizhe"). No prompt engineering required.

Install

npx tangerine-mcp           # zero-config, uses ~/.tangerine-memory

Or install globally:

npm install -g tangerine-mcp
tangerine-mcp --help

Usage

tangerine-mcp 0.1.0
MCP server that exposes Tangerine team memory to any MCP-compatible AI client.

USAGE
  npx tangerine-mcp [options]
  tangerine-mcp [options]

OPTIONS
  --root <path>     Memory root directory. Overrides $TANGERINE_MEMORY_ROOT.
                    Default: ~/.tangerine-memory
  -h, --help        Show this help and exit.
  -v, --version     Print version and exit.

ENVIRONMENT
  TANGERINE_MEMORY_ROOT   Memory root if --root is not given.

The server speaks MCP over stdio — stdout is JSONRPC, stderr is logs.

Add to Claude Code

Edit ~/.config/claude/config.json (macOS/Linux) or %APPDATA%\Claude\config.json (Windows):

{
  "mcpServers": {
    "tangerine": {
      "command": "npx",
      "args": ["-y", "tangerine-mcp"]
    }
  }
}

Restart Claude Code. The query_team_memory tool appears in the tool list. Ask Claude anything about prior decisions and it'll call the tool automatically.

Add to Cursor

Edit ~/.cursor/mcp.json:

{
  "mcpServers": {
    "tangerine": {
      "command": "npx",
      "args": ["-y", "tangerine-mcp"]
    }
  }
}

Add to Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "tangerine": {
      "command": "npx",
      "args": ["-y", "tangerine-mcp"]
    }
  }
}

Custom memory root

{
  "mcpServers": {
    "tangerine": {
      "command": "npx",
      "args": ["-y", "tangerine-mcp", "--root", "/path/to/team/memory"]
    }
  }
}

Or set the env var:

export TANGERINE_MEMORY_ROOT=/path/to/team/memory
npx tangerine-mcp

Memory layout

The server expects a directory of markdown files, optionally organized into the standard Tangerine subdirectories:

~/.tangerine-memory/
├── meetings/
│   └── 2026-04-25-v15-scope.md
├── decisions/
│   └── pricing-vercel-open-core.md
├── people/
├── projects/
├── threads/
└── glossary/

Each .md file may have YAML frontmatter:

---
title: Pricing Vercel-style open core (OSS forever + Cloud $10/seat + Enterprise)
date: 2026-04-26
source: meeting
---

We need to lock pricing before HN launch...

Frontmatter is parsed via gray-matter and surfaced in tool results.

Response envelope (Stage 1 AGI hook 4)

Every tool response is wrapped in this envelope — same shape, every tool, so MCP clients (Cursor, Claude Code, Claude Desktop) can render confidence indicators, freshness badges, and source attribution from day one. Stage 2 fills in the placeholder fields with real values.

{
  "data":              {"...": "actual tool payload — shape depends on the tool"},
  "confidence":        1.0,
  "freshness_seconds": 60,
  "source_atoms":      ["evt-2026-04-25-aBc12dEf", "..."],
  "alternatives":      [],
  "reasoning_notes":   null
}

| Field | Stage 1 (now) | Stage 2 (~6 mo out) | |---|---|---| | confidence | always 1.0 | LLM-graded 0..1 trust score | | freshness_seconds | seconds since the freshest source atom / file mtime | same | | source_atoms | atom ids that contributed ([] for substring search) | full citation graph | | alternatives | always [] | competing interpretations when ambiguous | | reasoning_notes | always null | reasoning loop annotation |

Schema is also documented in <root>/.tangerine/SCHEMA.md (written by the desktop daemon).

What it exposes

Tool: query_team_memory

Substring search across all memory markdown.

Input:

{
  "query": { "type": "string" },
  "limit": { "type": "number", "default": 5, "minimum": 1, "maximum": 20 }
}

data field shape:

{
  "query": "pricing",
  "root": "/path/to/memory",
  "searched": 12,
  "hits": [
    {
      "file": "decisions/pricing-vercel-open-core.md",
      "title": "Pricing Vercel-style open core (OSS forever + Cloud $10/seat + Enterprise)",
      "frontmatter": {"date": "2026-04-26", "source": "meeting"},
      "snippet": "...Vercel-style: OSS forever, Cloud $10/seat, Enterprise custom...",
      "content_preview": "<first 4000 chars>",
      "matches": 3
    }
  ]
}

Tool: get_today_brief()

Returns today's daily brief from <sidecar>/briefs/<today>.md if the desktop daemon has written it, otherwise synthesises one from the timeline index. data.origin is "brief_file" | "synthesised" | "empty".

Tool: get_my_pending(user)

Open action items where lifecycle.owner == user AND lifecycle.closed is null. Sorted overdue-first.

Tool: get_for_person(name)

Last 30 days, max 20 atoms involving the named person (matches refs.people and actors).

Tool: get_for_project(slug)

Last 30 days, max 20 atoms in the project (matches refs.projects).

Tool: get_thread_state(topic)

Chronological atoms attached to the thread, plus thread status (active | closed), decisions_resolved list, and the verbatim narrative from memory/threads/<topic>.md if present.

Tool: get_recent_decisions(days?)

Decision atoms (kind == "decision") in the last N days (default 7, max 365). Capped at 50 results, newest first.

Resource: team-memory://

• team-memory:// — JSON index of every file (path, title, frontmatter)
• team-memory://decisions/pricing-vercel-open-core.md — full markdown content with frontmatter

Behavior

• Memory root resolution: --root flag → $TANGERINE_MEMORY_ROOT → ~/.tangerine-memory
• Missing root: logs to stderr, returns empty results, doesn't crash
• Search: case-insensitive substring across body (frontmatter stripped), ranked by descending match count
• Caps: 1000 files, 4000-char content preview, 20 results max
• Performance: < 100ms for 100 files of ~5KB each
• Logging: stderr only (stdout reserved for MCP JSONRPC)

Development

git clone https://github.com/Tangerine-Intelligence/tangerine-meeting-live.git
cd tangerine-meeting-live/mcp-server
npm install
npm run build
npm test

License

Currently effective: Apache-2.0 — see PRIOR-LICENSE-APACHE-2.0.txt.

Transition draft (pending CEO + legal counsel ratification): AGPL v3 + Dual Commercial. See LICENSE and COMMERCIAL-LICENSE.md.