MCP Terminal Controller — macOS
Run commands, reuse persistent shells, follow long builds, and handle interactive prompts through the local mcp-terminal MCP server on macOS.
Free
Works with the AI Agents you already use
MCP Terminal Controller — macOS
Example session with this skill installed
Use MCP Terminal Controller in a temporary workspace to demonstrate persistent shell state. Set BUILD_MODE=debug, verify it in a second command, then run a three-step job with a short initial wait and follow its output without rerunning it. Handle a harmless project-name prompt by answering demo-project, verify each task's exit evidence, and close your session. Keep network access disabled.
- Read your context and instructions
- Compiled the mcp terminal controller
Verified local terminal demo
This result comes from a real macOS stdio MCP session with the workspace sandbox and approvals enabled. The harmless demo used no network access and created no background server.
| Check | Observed result |
|---|---|
| Persistent shell | Set BUILD_MODE in one command, read BUILD_MODE=debug in the next |
| Short commands | Both commands completed with exit code 0 |
| Long task | First call returned running after a 100 ms wait; the same session continued |
| Incremental output | Collected step 1, step 2, and step 3 without launching the task again |
| Completion evidence | The read returned last_command.exit_code=0 while the shell itself remained running |
| Interactive prompt | Detected awaiting_input, sent demo-project, and observed Selected project: demo-project |
| Final shell state | Idle before cleanup |
| Cleanup | terminal_close returned state=killed and no leftover process IDs |
Complete task output
BUILD_MODE=debug
step 1
step 2
step 3
Project name? demo-project
Selected project: demo-project
The task completed successfully, and its session was closed. For a real project, the agent uses the requested directory and commands, reports their actual exit evidence, and leaves a development server running only when continued operation is requested. The local MCP server must be installed and connected separately; loading this skill alone does not provide terminal access.
Connects securely to your tools. The creator never sees your data.
About this skill
MCP Terminal Controller is a free, open-source companion skill for MCP-server-controller, the macOS MCP server for sandboxed local terminal sessions.
It teaches an agent how to choose a one-shot command or a persistent shell, keep track of session IDs, follow long-running output without launching the command twice, answer authorized interactive prompts, and verify command completion using real exit evidence.
What it helps with
- Run short project commands with explicit working directories.
- Preserve shell state between dependent commands.
- Follow builds and tests with incremental output cursors.
- Start development servers and inspect readiness output.
- Work with interactive programs and rendered terminal screens.
- Respect server approvals, handle failures, and close owned sessions when finished.
Requirements
This is an instruction skill, not a hosted terminal service. The MCP-server-controller server must be installed separately on your Mac and connected to your AI client as a local stdio MCP server. Connecting only the Agensi MCP does not provide local terminal access.
Source/npm installation requires macOS and Node.js 22.13 or newer. The official Apple-silicon portable release includes its own runtime. Cloud-only clients without a connection to a local stdio server cannot execute these workflows.
Access and safety
The agent uses the terminal server to execute commands and access files allowed by your workspace, sandbox, and approval settings. It may use network access if permitted by your server configuration. The skill does not grant additional permissions and includes no telemetry or credential collection. Installing from source downloads dependencies; installation instructions identify the official repository and release assets.
Sandboxing, static command checks, and redaction are best-effort controls. The skill does not claim that every command or script is safe.
Included files
- SKILL.md: tool selection, execution, approvals, and reporting.
- references/setup.md: local installation and connection troubleshooting.
- references/workflows.md: concrete MCP call sequences.
- LICENSE: MIT license.
Project links
Repository: https://github.com/dogeheng/MCP-server-controller
Releases: https://github.com/dogeheng/MCP-server-controller/releases
Issues: https://github.com/dogeheng/MCP-server-controller/issues
macOS only. Requires a separately installed local stdio MCP server; Agensi's remote connector only delivers the skill instructions. Apple-silicon portable package is not for Intel Macs. Source/npm install requires Node.js 22.13+. Seatbelt depends on sandbox-exec. Sandbox writes are restricted, while file reads can be broader; command parsing and redaction are best-effort. PTY merges stdout/stderr. Interactive-input detection can be imperfect. Separate Claude Code, Cursor, and Copilot discovery flows have not been tested.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Free forever
No account required to browse
Trust & safety
Security scan
Security scan passed 20 days ago
- Free with an Agensi account
Requires
Permissions
Runs user-requested commands through the separately installed local MCP server. File writes follow configured workspace/sandbox limits; read access may be broader. Network access follows server policy and is used during installation or when the task needs it. Shell state may retain task environment variables; no credential harvesting or telemetry. Server approvals remain enabled.