mcp-cli

# MCP CLI Call any configured MCP server tool from the command line with `--flag=value` style args and full shell composition support. If `mcp-call` is not found, install it first: `uv tool install mcp-cli-skill` Always use `mcp-call` as the command, never `uvx mcp-cli-skill`. ## Commands ```bash mcp-call --search <words> # find tools across ALL servers, with params (start here) mcp-call --servers # list configured servers mcp-call <server> --tools # list tools (human-readable; --list / --help work too) mcp-call <server> --discover # list tools as JSON with schemas (= --tools -o json) # --servers / --tools / --search all take -o table|json|name; name prints `server` or `server tool` mcp-call <server> <tool> --schema # show tool's input schema as JSON mcp-call <server> <tool> --key=value ... # call a tool mcp-call <server> <tool> --input-json '{"key":"val"}' # call with JSON args echo '{}' | mcp-call <server> <tool> # call with stdin JSON mcp-call --add <name> <cmd> [args] [--env K=V ...] # add stdio server mcp-call --add-http <name> <url> [--header 'K: V' ...] # add HTTP server (auth via --header/-H) mcp-call --remove <name> # remove server mcp-call --login <name> [--no-browser] # OAuth browser login (the user runs this, not you) mcp-call --logout <name> # forget stored OAuth tokens mcp-call --sync # re-sync from Claude configs mcp-call --completion <bash|zsh|fish> # print shell completion hook mcp-call --refresh-completions # cache tool lists for tab completion mcp-call --clear-cache [server] # bust the completion cache ``` ## Server Management Config stored at `~/.mcp-cli/servers.json`. On first run, seeds from `~/.claude/settings.json` and `~/.claude.json`. Supports both stdio and HTTP MCP servers. ```bash mcp-call --add myredash uvx redash-mcp --env REDASH_URL=http://localhost --env REDASH_API_KEY=abc123 mcp-call --add github npx @modelcontextprotocol/server-github --env GITHUB_TOKEN=ghp_xxx mcp-call --remove myredash mcp-call --sync ``` Don't know which server has the tool you need? Run `mcp-call --search <words>` first: it searches every server's tool names and descriptions from a local cache in one call (name matches listed first, exit 1 if nothing matches), so you don't have to run `--tools` on each server. Each hit shows its params (`*` = required), so you can usually call the tool right away without `--help`. Read the default table and pick the server yourself: when the same tool is on several servers (e.g. `emergent-work` and `emergent-personal` are different accounts), never just take the first hit. `-o json` is there for jq. An unknown option or tool name exits 2 with a "Did you mean" hint and is never sent to the server — use the hint instead of guessing another name. If a call exits 4 with "needs interactive browser login", ask the user to run the `mcp-call --login <name>` command it prints; don't run it yourself. ## Examples ```bash mcp-call redash redash_query --action=adhoc --query="$(cat /tmp/query.sql)" --data_source_id=1 mcp-call redash redash_query --action=list --page_size=5 | jq '.results[].name' mcp-call slack slack_chat --action=post --channel=C123 --text="$(cat /tmp/msg.txt)" mcp-call redash redash_query --action=run --id=42 | jq '.query_result.data.rows' > /tmp/result.json ``` ## Multi-tool Workflow When you need to chain multiple MCP tools together, generate a bash script and run it: ```bash #!/bin/bash # Fetch github issues, search code, post to slack # 1. Fetch open bugs mcp-call github list_issues \ --owner=acme --repo=backend --state=open --labels=bug \ | jq '.[] | {number, title}' > /tmp/bugs.json # 2. Search for related code for title in $(jq -r '.[].title' /tmp/bugs.json | head -5); do mcp-call github search_code --query="$title repo:acme/backend" | jq '.items[:2]' done > /tmp/code_matches.txt # 3. Post summary to slack mcp-call slack send_message \ --channel="#engineering" \ --text="$(jq -r '.[] | "• #\(.number): \(.title)"' /tmp/bugs.json)" ``` This is better than making separate MCP tool calls because you get file I/O, pipes, loops, and variable expansion for free. ## When to Use Prefer this over MCP tool calls when: - Content comes from a file: `--arg="$(cat file)"` - Output needs piping: `| jq`, `> file`, `| grep` - Shell variable expansion needed: `--arg="$VAR"` - Chaining multiple calls in one command ## Arg Types Values auto-parse: `--id=42` → int, `--flag=true` → bool, `--name=hello` → string. ## Colliding param names If a tool has a param named `help`, `schema`, or `input-json` (which clash with mcp-call's own flags), pass it after `--`: ```bash mcp-call srv tool --id=5 -- --help=true --schema=foo # everything after -- goes to the tool verbatim ```