Run a command through MCP
github.com/libtmux/libtmux-go/mcp · Source
This program creates a private session, selects its active pane, and exposes
only run_shell_command to the MCP client.
The shell command prints one line and deliberately exits with status 7.
The client checks completion and output before reporting the result.
Run the example¶
Use Go 1.26 or newer and tmux 3.2a or newer on Linux, macOS, or WSL. The project pins the core library and MCP executable to their published releases. The client uses the official Go MCP SDK.
Create an empty directory:
$ mkdir run-mcp-commandEnter it:
$ cd run-mcp-commandInstall the MCP executable inside the project:
$ GOBIN="$PWD/.tools" go install \ github.com/libtmux/libtmux-go/mcp/cmd/libtmux-mcp@v0.0.1-alpha.12Save the module file:
module example.com/command-runner
go 1.26.0
require ( github.com/libtmux/libtmux-go v0.0.1-alpha.9 github.com/modelcontextprotocol/go-sdk v1.6.1)
require ( github.com/google/jsonschema-go v0.4.3 // indirect github.com/segmentio/asm v1.1.3 // indirect github.com/segmentio/encoding v0.5.4 // indirect github.com/yosida95/uritemplate/v3 v3.0.2 // indirect golang.org/x/oauth2 v0.35.0 // indirect golang.org/x/sys v0.41.0 // indirect)Client program¶
Save this as main.go:
package main
import ( "context" "encoding/json" "errors" "fmt" "os" "os/exec" "path/filepath" "time"
"github.com/libtmux/libtmux-go/tmux" sdk "github.com/modelcontextprotocol/go-sdk/mcp")
func main() { if err := run(); err != nil { fmt.Fprintln(os.Stderr, err) os.Exit(1) }}
func run() (result error) { directory, err := os.MkdirTemp("", "libtmux-go-command-") if err != nil { return err }
server, err := tmux.NewServer(tmux.ServerOptions{ SocketPath: filepath.Join(directory, "tmux.sock"), ConfigFile: "/dev/null", }) if err != nil { return errors.Join(err, os.RemoveAll(directory)) } defer func() { cleanup, cancel := context.WithTimeout(context.Background(), 5*time.Second) defer cancel() if err := server.Kill(cleanup); err != nil { result = errors.Join(result, fmt.Errorf( "stop tmux; resources retained at %s: %w", directory, err, )) return } result = errors.Join(result, os.RemoveAll(directory)) }()
ctx, cancel := context.WithTimeout(context.Background(), 20*time.Second) defer cancel() created, err := server.NewSession(ctx, tmux.NewSessionRequest{ Name: "demo", Command: "sh", }) if err != nil { return err }
pane, err := created.ResolveActivePane(ctx) if err != nil { return err }
binary, err := filepath.Abs(".tools/libtmux-mcp") if err != nil { return err } command := exec.Command(binary, "-socket-path", server.SocketPath(), "-binary", server.Executable(), ) command.Stderr = os.Stderr command.Env = []string{ "PATH=" + os.Getenv("PATH"), "HOME=" + os.Getenv("HOME"), "LANG=C.UTF-8", "TMPDIR=" + directory, "LIBTMUX_TOOLSETS=", "LIBTMUX_TOOLS=run_shell_command", } client := sdk.NewClient(&sdk.Implementation{ Name: "command-runner", Version: "1.0.0", }, nil) connection, err := client.Connect(ctx, &sdk.CommandTransport{ Command: command, TerminateDuration: 3 * time.Second, }, nil) if err != nil { return fmt.Errorf("connect MCP client: %w", err) } defer func() { if err := connection.Close(); err != nil { result = errors.Join(result, fmt.Errorf("close MCP client: %w", err)) } }()
reply, err := connection.CallTool(ctx, &sdk.CallToolParams{ Name: "run_shell_command", Arguments: map[string]any{ "pane_id": string(pane.ID()), "command": "printf 'command finished\\n'; (exit 7)", "timeout": 5, "max_lines": 20, }, }) if err != nil { return fmt.Errorf("call run_shell_command: %w", err) } if reply.IsError { for _, content := range reply.Content { if text, ok := content.(*sdk.TextContent); ok { return fmt.Errorf("run_shell_command: %s", text.Text) } } return errors.New("run_shell_command failed without a text explanation") }
data, err := json.Marshal(reply.StructuredContent) if err != nil { return err } var ran struct { PaneID string `json:"pane_id"` ResolvedPaneIDs []string `json:"resolved_pane_ids"` ExitStatus *int `json:"exit_status"` TimedOut bool `json:"timed_out"` Output []string `json:"output"` OutputUnavailable string `json:"output_unavailable"` LinesMissed bool `json:"lines_missed"` } if err := json.Unmarshal(data, &ran); err != nil { return fmt.Errorf("decode command result: %w", err) } if ran.TimedOut || ran.ExitStatus == nil { return errors.New("command completion is unconfirmed; do not retry it automatically") } if ran.PaneID != string(pane.ID()) || len(ran.ResolvedPaneIDs) != 1 || ran.ResolvedPaneIDs[0] != ran.PaneID { return fmt.Errorf("unexpected pane selection: %s", data) } if ran.OutputUnavailable != "" || ran.LinesMissed { return fmt.Errorf("command exited %d, but its output is incomplete: %s", *ran.ExitStatus, data) } fmt.Printf("exit status: %d\n", *ran.ExitStatus) for _, line := range ran.Output { fmt.Println(line) } return nil}Resolve the module dependencies:
$ go mod tidyRun the program from the directory containing both files and .tools:
$ go run .Its stdout is:
exit status: 7command finishedCheck completion and output¶
Session.ResolveActivePane supplies the pane ID. The request uses the public
wire fields pane_id, command, timeout, and max_lines; the client checks
that the reply names the same pane exactly once.
A nonzero shell exit status is distinct from a tool error. The request here
succeeds, completion is observed, and the shell’s status is 7. A missing exit
status or timed_out means completion is unconfirmed. The program reports an
error instead of automatically sending the command again.
It also rejects output_unavailable and lines_missed. A command can finish
successfully while its terminal output is incomplete. The request retains at
most the newest 20 lines. The public reply does not report whether that cap
discarded earlier lines. Applications that retain partial output should display
its bounds and limitations alongside it. See
Waits and output for timeout behavior and
other observation tools.
The public command handler validates the request and constructs the snake_case reply. The client decodes that reply rather than depending on an internal Go result type.
Shutdown¶
Startup and requests share a 20-second context. During shutdown, the client closes the child’s stdin, then escalates to SIGTERM and SIGKILL if needed. It waits up to three seconds after each step. Deferred cleanup closes the client before stopping the owned tmux server with a fresh five-second context, so an expired request context does not skip daemon cleanup.
The child receives a curated environment with the selected socket and tmux
binary. Its temporary files live under the program’s owned directory through
TMPDIR. It does not inherit TMUX, TMUX_PANE, or unrelated LIBTMUX_*
settings from your interactive shell.
The program joins operation and cleanup errors. It removes its temporary
directory only after Server.Kill succeeds; otherwise, it prints the retained
directory for inspection. Its server cleanup implementation
defines the library’s kill behavior.
The connection guide covers connecting to an existing server instead. The example’s cleanup is appropriate because it creates its own private socket and session.