The mcp bundle ships Model Context Protocol servers — small processes (or remote endpoints) that extend any MCP-aware AI harness with new capabilities. lfg installs the server and wires it into every supported harness present on your $PATH.

The bundle sits directly below skills in the tree picker, gated on the same node-lts requirement (npm-distributed servers are still the most common shape).

What lfg does

For each MCP server you select:

1. Install — for stdio servers, runs npm install -g <package> so the first launch doesn't pay the npx download cost. For remote (HTTP / SSE) servers, this step is a no-op.
2. Prompt for credentials — between confirm and progress, lfg shows a credentials wizard: one masked input per unique env var declared by the selected MCPs. Inputs are deduped (a token shared by two servers is asked once). Leave any field blank to skip — you can export FOO=... in your shell rc later; the done screen reminds you of every blank.
3. Register — for every harness in the entry's target_harnesses list that's actually on $PATH, lfg invokes the harness's own mcp add command (or merges into its config file when there is no CLI helper). Collected credentials are baked into the harness config: stdio entries get --env KEY=VAL, remote entries get ${VAR} placeholders in URLs/headers expanded inline.
4. Surface unset env vars — any var the user left blank in the wizard is shown on the done screen as a copy-paste export FOO=... reminder.

Supported harnesses

Harness	Binary	How lfg registers	Storage
Claude Code	claude	claude mcp add --scope user	~/.claude.json
Codex	codex	codex mcp add	~/.codex/config.toml
Droid	droid	droid mcp add	~/.factory/mcp.json
opencode	opencode	merge into config JSON	~/.config/opencode/opencode.json

Adding a new harness to lfg means dropping a harnessAdapter into internal/installer/mcp_register.go and adding it to the init() list — every existing MCP entry then auto-extends to the new harness on the next run (target_harnesses = ["all"] is the recommended default for built-in entries).

Schema

MCP entries live under any bundle (the default preset puts them in mcp). Every field is documented in the Preset TOML schema; the MCP-specific subset is:

[[bundles.tools]]
source       = "mcp"           # routes to the mcp installer
name         = "my-server"     # display name; "mcp-" prefix stripped on register

# --- one of: stdio (default) | http | sse ---
mcp_type     = "stdio"

# --- stdio fields ---
mcp_package  = "@scope/server-name"   # npm package to install globally
mcp_command  = ""                     # optional: override launch binary (default: npx)
mcp_args     = []                     # extra args appended after the package

# --- remote (http / sse) fields ---
mcp_url      = "https://mcp.example.com/sse"
mcp_headers  = { "Authorization" = "Bearer ..." }   # sent on every request

# --- common ---
target_harnesses = ["all"]            # or e.g. ["claude-code", "codex"]
env_vars         = ["MY_SECRET"]      # surfaced in info card + done screen

The target_harnesses = ["all"] shorthand expands at install time to every harness lfg knows about, so a custom preset stays in sync as new harnesses are added.

Custom preset example

Add a private remote MCP and ship it to every installed harness:

[[bundles]]
id   = "mcp"
name = "mcp"
description = "Custom MCP servers"

[[bundles.tools]]
name             = "internal-billing"
source           = "mcp"
description      = "Read internal billing tool"
mcp_type         = "http"
mcp_url          = "https://mcp.internal.corp/billing"
mcp_headers      = { "Authorization" = "Bearer $BILLING_TOKEN" }
target_harnesses = ["all"]
env_vars         = ["BILLING_TOKEN"]

Run lfg --config ./preset.toml apply and the entry registers with every harness on your machine; BILLING_TOKEN shows up on the done screen as a reminder to export it.

Troubleshooting

• "skipping registration: claude not on PATH" — the harness isn't installed. Install it (it's in the ai-harnesses bundle) and re-run lfg apply; only the missing registrations re-execute.
• "warning: claude-code registration failed — exit 1" — a harness CLI rejected the add. The literal command lfg ran is in the log tail; copy it, run by hand, and the error message tells you what to fix (auth, conflicting name, malformed URL).
• Set the env var first, then the MCP server connects — harnesses inherit env from the parent shell at launch, not from the registered config. After exporting in your rc, restart the harness.