Karbon MCP Server
An unofficial Model Context Protocol server that connects MCP-compatible clients to Karbon practice management, enabling natural language interaction with contacts, work items, notes, timesheets, and more.
README
Karbon MCP Server
🚧 Early release — this project is brand new and we're actively making updates over the next few days. Things may change quickly; pin a version if you need stability, and check back soon.
An unofficial Model Context Protocol server for Karbon practice management. Connect Claude Code, Claude Desktop, GitHub Copilot, or any MCP-compatible client to your Karbon instance and work with contacts, work items, notes, timesheets, and more using natural language.
Disclaimer: This is a community project and is not affiliated with or endorsed by Karbon. Use at your own risk.
Features
| Area | Tools |
|---|---|
| Contacts | list_contacts, get_contact, create_contact, update_contact |
| Organizations | list_organizations, get_organization, create_organization, update_organization |
| Work items | list_work_items, get_work_item, create_work_item, update_work_item |
| Notes | get_note, create_note |
| Timesheets | list_timesheets, get_timesheet, list_time_entries (read-only) |
| Users | list_users, get_user |
| Account | get_tenant_settings, list_work_templates |
All list tools support OData filtering ($filter), ordering ($orderby), and pagination ($top/$skip) — so you can ask things like "show my in-progress work items due this month" and the AI can express that as a precise query.
Prerequisites
- Node.js 18+
- Karbon API credentials. In Karbon go to Settings → Connected Apps → API Applications and copy your Bearer Token and Access Key. (Karbon help article)
Setup
Step-by-step walkthroughs for Claude Desktop, claude.ai/Cowork, and ChatGPT (written for non-developers) are in docs/client-guides.md.
One-line install (nothing to download first)
No Node.js? These check for it (installing via winget/Homebrew where possible), then run the setup wizard straight from npm — no repo clone, no manual download:
# Windows (PowerShell)
irm https://raw.githubusercontent.com/Mad-Man-Dan/karbon-mcp-server/main/scripts/install.ps1 | iex
# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/Mad-Man-Dan/karbon-mcp-server/main/scripts/install.sh | bash
Quickest with Node.js installed: automatic setup
npx -y karbon-mcp-server setup
Setup asks a few quick questions — project or user-level install, your API keys (skippable), and whether to enable read-only mode — then does everything else automatically:
- This project (the default): writes
.mcp.json(Claude Code), plus.vscode/mcp.jsonand.cursor/mcp.jsonif those folders exist in the project. - User account: detects which MCP clients are installed on your machine — Claude Code, Codex, Claude Desktop, VS Code, Cursor — and adds the server to each (for Claude Code and Codex it uses their own
mcp addCLI).
It then offers to take your two Karbon keys (Enter skips either one):
- Input is masked — you see a
*per character, then a first-4/last-4 preview (like1a2b****5c6d, the same style Karbon's site uses) so you can confirm what was entered, with a gentle warning if the format looks off (the Bearer Token is a GUID; the Access Key starts withey...). - Keys are written only to your MCP client's local config file on your machine — the same file a manual setup would use. They are never sent anywhere else.
- If you skip, placeholder values are written and setup tells you exactly which file(s) to edit later.
It backs up every existing config file before touching it, merges rather than overwrites, and never replaces real credentials with placeholders (re-running and skipping the prompts keeps your existing keys). If you go project-level, gitignore the config files — your keys live in them in plain text.
Options:
npx -y karbon-mcp-server setup --scope project # skip the question
npx -y karbon-mcp-server setup --scope user
npx -y karbon-mcp-server setup --read-only # disable all write tools
npx -y karbon-mcp-server setup --client vscode,cursor # only these clients
# ids: claude-code, codex, claude-desktop, vscode, cursor
npx -y karbon-mcp-server setup --print # just print the JSON
One-click installs
The Install in VS Code / Cursor badges above create the server entry for you with placeholder credentials. After installing, replace YOUR_BEARER_TOKEN and YOUR_ACCESS_KEY with your real keys — in VS Code via the Command Palette → MCP: Open User Configuration, in Cursor by editing ~/.cursor/mcp.json.
If you run the server before replacing the placeholders, it starts fine and every tool responds with setup instructions instead of failing cryptically — so you can even ask your AI assistant "why isn't Karbon working?" and it will tell you what to do.
Prefer to configure things by hand? Each client is below.
Claude Code
claude mcp add karbon \
--env KARBON_BEARER_TOKEN=your-bearer-token \
--env KARBON_ACCESS_KEY=your-access-key \
-- npx -y karbon-mcp-server
Claude Desktop
Add to claude_desktop_config.json:
{
"mcpServers": {
"karbon": {
"command": "npx",
"args": ["-y", "karbon-mcp-server"],
"env": {
"KARBON_BEARER_TOKEN": "your-bearer-token",
"KARBON_ACCESS_KEY": "your-access-key"
}
}
}
}
VS Code (GitHub Copilot)
Add to .vscode/mcp.json in your workspace (or your user mcp.json):
{
"servers": {
"karbon": {
"command": "npx",
"args": ["-y", "karbon-mcp-server"],
"env": {
"KARBON_BEARER_TOKEN": "your-bearer-token",
"KARBON_ACCESS_KEY": "your-access-key"
}
}
}
}
Cursor
Add to ~/.cursor/mcp.json (or use the install badge above):
{
"mcpServers": {
"karbon": {
"command": "npx",
"args": ["-y", "karbon-mcp-server"],
"env": {
"KARBON_BEARER_TOKEN": "your-bearer-token",
"KARBON_ACCESS_KEY": "your-access-key"
}
}
}
}
Codex (OpenAI Codex CLI)
codex mcp add karbon \
--env KARBON_BEARER_TOKEN=your-bearer-token \
--env KARBON_ACCESS_KEY=your-access-key \
-- npx -y karbon-mcp-server
Or add to ~/.codex/config.toml:
[mcp_servers.karbon]
command = "npx"
args = ["-y", "karbon-mcp-server"]
[mcp_servers.karbon.env]
KARBON_BEARER_TOKEN = "your-bearer-token"
KARBON_ACCESS_KEY = "your-access-key"
ChatGPT, claude.ai chat & Claude Cowork (remote connectors)
These clients can't launch local servers — they only connect to a remote MCP server over HTTPS. This package includes a streamable-HTTP mode for exactly that, so a firm can self-host it and keep their Karbon keys in their own infrastructure:
KARBON_BEARER_TOKEN=... KARBON_ACCESS_KEY=... KARBON_HTTP_SECRET=some-long-random-string \
npx -y karbon-mcp-server --http --port 8787
This serves the MCP endpoint at /mcp/<your-secret> (plus a /health check). Expose it over HTTPS however you like — the quickest free option is a Cloudflare Tunnel:
cloudflared tunnel --url http://localhost:8787
Then add the URL (e.g. https://your-tunnel.trycloudflare.com/mcp/your-secret) as:
- claude.ai / Cowork: Settings → Connectors → Add custom connector (Pro/Max/Team/Enterprise plans)
- ChatGPT: Settings → Connectors (requires developer mode for full tool access)
Security notes: anyone with the full URL has the same Karbon access as your API keys, so use a long random
KARBON_HTTP_SECRET, always front it with HTTPS, and considerKARBON_READ_ONLY=truefor remote use. There is deliberately no shared/hosted instance of this server — your keys should only ever live on infrastructure you control.
Configuration
| Environment variable | Required | Description |
|---|---|---|
KARBON_BEARER_TOKEN |
Yes | Bearer token from Karbon Connected Apps |
KARBON_ACCESS_KEY |
Yes | Access key (JWT) from Karbon Connected Apps |
KARBON_READ_ONLY |
No | Set to true to disable all create/update tools |
KARBON_API_BASE_URL |
No | Override the API base URL (default https://api.karbonhq.com/v3) |
KARBON_HTTP_SECRET |
No | HTTP mode only: secret path segment for the endpoint (/mcp/<secret>) |
PORT |
No | HTTP mode only: port to listen on (default 8787; --port wins) |
Read-only mode
If you want a guarantee that the AI can never modify your Karbon data, set KARBON_READ_ONLY=true (or pass --read-only). Write tools are not registered at all in this mode, so the AI never sees them. The setup wizard asks about this.
Changing it later: setup always writes the toggle explicitly ("KARBON_READ_ONLY": "true" or "false"), so the switch is right there in your config file — flip the value and restart the client. Write tools appear or disappear accordingly (in read-only mode they aren't registered at all, so the AI never sees them). Alternatively, re-run setup and skip the key prompts (press Enter) — your saved keys are kept, and your new read-only answer is applied. One caveat: entries managed by the claude / codex CLIs aren't touched by a re-run — for those, run claude mcp remove karbon (or codex mcp remove karbon) first, then re-run setup.
A note on credentials
Whether you type your keys into the setup wizard or paste them into a config file by hand, they end up in exactly one place: your own MCP client's local config file — the same file a fully manual setup would use. Nothing in this project transmits your keys anywhere else. Like every stdio MCP server, the keys live in that file in plain text, so treat it like a password: it grants the same access to your Karbon account as your keys do. If a key is ever exposed, regenerate it in Karbon under Settings → Connected Apps.
Example prompts
- "Find the contact record for Jane Smith and show her business card details."
- "List all in-progress work items assigned to me, ordered by due date."
- "Create a work item for Acme Corp's 2025 tax return, assigned to alex@firm.com, starting next Monday."
- "Add a note to the Smith work item summarizing today's client call."
- "How many hours were logged against client work last week?"
Development
git clone https://github.com/Mad-Man-Dan/karbon-mcp-server.git
cd karbon-mcp-server
npm install
npm run build
Test locally with the MCP Inspector:
npx @modelcontextprotocol/inspector \
-e KARBON_BEARER_TOKEN=... -e KARBON_ACCESS_KEY=... \
node dist/index.js
The Karbon API is documented at developers.karbonhq.com with an OpenAPI spec at karbonhq/karbon-api-reference.
Roadmap
- [ ] Client groups, invoices, and custom fields tools
- [ ] Files and comments
- [ ] One-click Claude Desktop extension (
.mcpbbundle) - [x] Remote mode (streamable HTTP) for ChatGPT / claude.ai / Cowork
- [ ] One-click "Deploy to Cloudflare Workers" template for the remote mode
- [ ] Publish to the MCP registry
Contributions welcome — open an issue or PR.
License
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.