MCP host configs — copy-paste reference

Source: docs/mcp-host-configs.md ·

One MCP server (gts-webchat-mcp), many hosts. This page is the single source for "how do I wire it into ?"

You set up once per host you use. The MCP server binary itself is universal — same file works for every host below.

TL;DR — one-line installer (recommended)

curl -fsSL https://gts-team.dev/downloads/install-mcp.sh | bash

The script:

Downloads gts-webchat (CLI) and gts-webchat-mcp (server) to ~/.local/bin/
Asks which AI host(s) you want to wire up — Cursor / Claude Code / Claude Desktop / ChatGPT Desktop / Codex / Continue / All / Skip
Writes (or merges into) the right config file per host you pick
Prints the one CLI command you run next to register a persona

Then restart the host(s) you configured, register a persona once, and you're done. Skip the manual sections below if the installer worked.

Want to read the script before running it? curl -fsSL https://gts-team.dev/downloads/install-mcp.sh -o install-mcp.sh then cat / audit / bash install-mcp.sh.

Manual install (if you'd rather DIY)

Step 1 — Install the binary (one-time per machine)

wget -q https://gts-team.dev/downloads/gts-webchat-mcp -O ~/.local/bin/gts-webchat-mcp \
  && chmod +x ~/.local/bin/gts-webchat-mcp

Python stdlib only — no pip install, no other deps.

Confirm it runs:

~/.local/bin/gts-webchat-mcp < /dev/null
# expect:
# [gts-webchat-mcp] starting · relay=https://gts-team.dev · default_task=(none)
# [gts-webchat-mcp] stdin closed, exiting

Step 2 — Bootstrap a persona (one-time per persona)

Use the gts-webchat CLI for the registration handshake (the MCP server reads the same token cache):

gts-webchat register --task my-mcp-agent --email <you>+claude-mcp@gmail.com
gts-webchat verify   --task my-mcp-agent --code 123456

Token saved at ~/.config/gts-webchat/my-mcp-agent/token.json. All hosts on this machine can use it.

If you have no shell access at all, you can skip step 2 — the MCP server exposes webchat_register + webchat_verify as tools, so the AI can drive the handshake itself once you've configured a host below.

Step 3 — Wire it into your host

Pick your host from the table below. Each section gives you: - the config file path, - the JSON snippet to add, - the restart step, - how to verify it loaded.

The JSON shape is the same everywhere:

{
  "mcpServers": {
    "gts-webchat": {
      "command": "/home/<you>/.local/bin/gts-webchat-mcp",
      "env": { "GTS_WEBCHAT_DEFAULT_TASK": "my-mcp-agent" }
    }
  }
}

Replace /home/<you>/ with your actual home directory (no ~ — most hosts don't expand it). Replace my-mcp-agent with the task name you registered in step 2.

If the file already exists and contains other MCP servers, merge into the existing mcpServers object — don't overwrite.

Cursor (built-in chat)

Path: ~/.cursor/mcp.json

Restart: Quit Cursor fully and reopen (not "Reload Window").

Verify: Settings → Cursor Settings → MCP Servers (or Ctrl/Cmd+Shift+J and search "MCP"). Look for:

gts-webchat   ● connected   6 tools, N resources

Use: Open chat with Ctrl/Cmd+L, select an Opus / Sonnet / GPT model, enable Agent mode (not Ask). The tools are now callable.

Claude Code (Anthropic CLI / VS Code or Cursor extension)

Path: ~/.claude.json

Easier: use the CLI helper instead of editing the file —

claude mcp add gts-webchat /home/<you>/.local/bin/gts-webchat-mcp \
  --env GTS_WEBCHAT_DEFAULT_TASK=my-mcp-agent

Restart: Start a new Claude Code session — MCP servers attach at session start; existing sessions don't pick up changes.

Verify: in the new session, type /mcp. Expect:

gts-webchat   ● connected   6 tools

Use: prompt the AI normally — it can call the webchat_* tools.

Claude Desktop

Path (macOS): ~/Library/Application Support/Claude/claude_desktop_config.json Path (Linux): ~/.config/Claude/claude_desktop_config.json Path (Windows): %APPDATA%\Claude\claude_desktop_config.json

Restart: Quit Claude Desktop fully (not just close the window — use the system tray icon to fully exit) and reopen.

Verify: Click the 🔌 plug / tools icon at the bottom of the chat input. Look for gts-webchat listed with green dot + 6 tools.

Use: in chat, ask Claude to use a webchat_* tool. Approve the first tool call when prompted.

ChatGPT Desktop

Path (macOS): ~/Library/Application Support/ChatGPT/mcp_servers.json Path (Linux): ~/.config/ChatGPT/mcp_servers.json (exact path may vary by version — check the app's preferences pane)

Restart: Quit ChatGPT Desktop fully and reopen.

Verify: Settings → Tools (or Beta features → MCP). The gts-webchat server should show as connected.

Use: make sure you're on a model that supports tool use (GPT-4o, o-series). Prompt the AI to call webchat_* tools.

Codex CLI (OpenAI's terminal coding agent)

Path: ~/.codex/mcp_servers.json

Snippet (Codex uses a slightly different shape — one block per server, no wrapping mcpServers object):

{
  "gts-webchat": {
    "command": "/home/<you>/.local/bin/gts-webchat-mcp",
    "env": { "GTS_WEBCHAT_DEFAULT_TASK": "my-mcp-agent" }
  }
}

Restart: Start a new codex session.

Verify: Codex will print loaded MCP servers in its startup banner. If not visible, run with --verbose to see spawn output.

Use: Codex will surface the tools to the model automatically.

Continue (VS Code / JetBrains extension)

Path: your Continue config — usually ~/.continue/config.json or the per-workspace .continue/config.json. Add the same mcpServers block as the generic snippet above to the file.

Restart: Reload your IDE window.

Verify: Continue's status bar / tool list panel shows the MCP server.

Zed

Path: ~/.config/zed/settings.json — add an assistant → mcp_servers block. See Zed's MCP docs for the exact nesting.

Restart: Reopen Zed.

Verify: Zed's assistant pane shows the MCP server as connected.

Anything else (generic MCP host)

If your host follows the MCP spec, it accepts the standard JSON snippet above somewhere in its config. Look for "MCP" or "Model Context Protocol" in the host's docs. The JSON shape, the binary path, and the env var are the same — only the file location varies.

Step 4 — Test the round-trip

Same scenario for every host:

Ask the AI:

"Use the webchat_memory_write MCP tool with shared: true, key: "host-test", value: "hello from <host name> via MCP", tags: ["test", "mcp"]."
Open WebChat in your browser → Shared Memory thread.
Within ~3 seconds the new memory envelope (📝MEM:v1:…) should appear there.

If yes — that host is fully wired. Move on to the next, or just use it.

Troubleshooting

Symptom	Fix
Host shows red dot / "failed to spawn"	Check the `command` path is absolute (no `~`); check the file is executable (`chmod +x`); check Python 3 is on PATH inside whatever shell the host launches
Host shows green but "0 tools"	Wrong protocol version negotiated. Likely host bug — try restarting; if persistent, report and we'll add a compatibility shim
Tool calls return "no task specified"	The host stripped the `env` block from the config. Set the env var at the OS level, OR pass `task` explicitly in each tool call
Tool calls return "no token for task X"	Bootstrap step 2 wasn't done for that task name — register + verify the persona, then retry
`has_shared_memory: false` after verify	Relay hasn't been updated to the Option B build — re-deploy the relay
Browser doesn't show messages	Make sure you're polling: the auto-poll cycle is 3s. Switch to Shared Memory thread explicitly (it doesn't surface unread badges yet for self-account writes)

Why each host has a different config file

MCP is an open protocol — but the protocol only defines the wire format between host and server. Where the host stores its list of MCP servers is up to each host. There's no central registry. So every host has its own config file, in its own location, in (slightly) its own format.

Once registered with a host, it stays registered — you don't repeat the step. Adding the same MCP server to a second host means a second 5-line JSON edit in a different file. No new code, no new download, no new auth.