>smartsh_

Safe, compact command execution for AI coding agents.

MCP server + local daemon that gives Cursor and Claude Code a safe, token-efficient way to run shell commands.

Why smartsh?

When AI agents run terminal commands, they dump huge raw logs into context — burning tokens and confusing the model.

smartsh fixes this:

• Runs commands through a local daemon (smartshd)
• Returns compact structured JSON instead of raw output
• Uses Ollama for intelligent summaries of command results
• Applies safety checks before execution (blocks dangerous commands)
• Supports risk approval workflows for destructive operations
• Truncates output automatically for massive token savings

Install (3 steps)

Step 1: Install Ollama (required)

smartsh uses Ollama for intelligent command output summaries. Install it first.

macOS / Linux:

curl -fsSL https://ollama.com/install.sh | sh
ollama serve
ollama pull llama3.2:3b

Windows:

1. Download and install from https://ollama.com/download
2. Then in PowerShell:

ollama serve
ollama pull llama3.2:3b

Step 2: Install smartsh

If Ollama is already installed, this is the only terminal command users need to run for smartsh installation.

macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/BegaDeveloper/smartsh/main/scripts/install.sh | sh

Windows (PowerShell):

powershell -NoProfile -ExecutionPolicy Bypass -Command "iwr -useb https://raw.githubusercontent.com/BegaDeveloper/smartsh/main/scripts/install.ps1 | iex"

Install via Go:

go install github.com/BegaDeveloper/smartsh/cmd/smartsh@latest
go install github.com/BegaDeveloper/smartsh/cmd/smartshd@latest
smartsh setup-agent

The installer runs setup-agent automatically. After setup completes, you get these files in ~/.smartsh/:

Step 3: Connect to Cursor or Claude Code

For Cursor:

Copy cursor-mcp.json into your project:

# macOS / Linux
cp ~/.smartsh/cursor-mcp.json /path/to/your/project/.cursor/mcp.json

# Windows
copy %USERPROFILE%\.smartsh\cursor-mcp.json C:\path\to\your\project\.cursor\mcp.json

Then paste ~/.smartsh/agent-instructions.txt into Cursor → Settings → Rules.

For Claude Code:

Copy claude-code-mcp.json as your Claude desktop config:

# macOS / Linux
cp ~/.smartsh/claude-code-mcp.json ~/.claude/claude_desktop_config.json

# Windows
copy %USERPROFILE%\.smartsh\claude-code-mcp.json %USERPROFILE%\.claude\claude_desktop_config.json

Verify

smartsh doctor

Generated MCP JSON (reference)

If you prefer to set up manually, here's what the generated cursor-mcp.json / claude-code-mcp.json looks like:

If you are writing JSON manually, use your own absolute smartsh binary path for command (it differs per user and OS).

{
  "mcpServers": {
    "smartsh": {
      "command": "/ABSOLUTE/PATH/TO/smartsh",
      "args": ["mcp"],
      "env": {
        "SMARTSH_DAEMON_URL": "http://127.0.0.1:8787",
        "SMARTSH_DAEMON_TOKEN": "<auto-generated-token>",
        "SMARTSH_SUMMARY_PROVIDER": "ollama",
        "SMARTSH_OLLAMA_REQUIRED": "true",
        "SMARTSH_OLLAMA_ALWAYS": "true",
        "SMARTSH_OLLAMA_URL": "http://127.0.0.1:11434",
        "SMARTSH_OLLAMA_MODEL": "llama3.2:3b",
        "SMARTSH_MCP_HTTP_TIMEOUT_SEC": "900",
        "SMARTSH_MCP_DEFAULT_UNSAFE": "true",
        "SMARTSH_MCP_DEFAULT_REQUIRE_APPROVAL": "false",
        "SMARTSH_MCP_DEFAULT_ALLOWLIST_MODE": "off"
      }
    }
  }
}

The token is generated automatically by smartsh setup-agent and stored in ~/.smartsh/config.

Rule snippet (paste into Cursor/Claude rules)

For command execution, always use the smartsh MCP tool (smartsh_run).
In some clients it appears as smartsh-local_smartsh_run.
Never use direct shell execution unless explicitly requested by the user.
Do not use run_terminal_cmd when the smartsh MCP tool is available.
Prefer summarized tool output and avoid dumping full terminal logs.

How It Works

┌─────────────┐     MCP JSON-RPC      ┌─────────┐     HTTP      ┌──────────┐
│ Cursor /     │ ──────────────────▶  │ smartsh  │ ──────────▶  │ smartshd │
│ Claude Code  │ ◀──────────────────  │   mcp    │ ◀──────────  │  daemon  │
└─────────────┘   compact summary     └─────────┘   execute     └──────────┘

1. Agent sends command via MCP tool (smartsh_run)
2. smartsh mcp forwards to local daemon
3. smartshd validates safety → executes → summarizes via Ollama
4. Compact JSON returned to agent (not raw logs)

Example response

{
  "status": "failed",
  "exit_code": 1,
  "summary": "command failed (exit code 1): Cannot find module '@app/auth'",
  "error_type": "compile",
  "primary_error": "Cannot find module '@app/auth'",
  "next_action": "Fix TypeScript compiler errors and rerun build/test.",
  "failed_files": ["src/app/auth/auth.service.ts"],
  "top_issues": ["TS2307: Cannot find module '@app/auth'"]
}

Compare that to 500+ lines of raw tsc output the agent would normally dump.

Features

Safety & Policy

• Blocks dangerous commands (rm -rf /, privilege escalation, pipe-to-shell)
• Risk approval workflow — agent must confirm before running destructive ops
• Command allowlist mode (off / warn / enforce)
• Project-level policy via .smartsh-policy.yaml

Token Savings

• Success runs return only summary (no output tail)
• Failed runs return truncated tail + structured error info
• MCP compact mode enabled by default
• Configurable tail size via SMARTSH_MCP_MAX_OUTPUT_TAIL_CHARS

Ollama Summaries

• Ollama is the default summary provider
• Sends only truncated, redacted output to local Ollama
• Strict JSON response schema enforced
• Falls back to deterministic parsing if Ollama is unavailable

Daemon Capabilities

• Persistent jobs in BoltDB (survive restarts)
• Async execution with job_id polling
• SSE status streaming
• PTY interactive sessions
• Execution isolation (timeout, memory, CPU, env allowlist)
• Token auth required by default
• Prometheus metrics at /metrics

Configuration

Environment Variables

Risky Commands

When an agent tries to run a destructive command (e.g. rm -rf), smartsh returns status=needs_approval with an approval_id. The agent then calls smartsh_approve with decision=yes or decision=no.

Use unsafe=true in the tool call only when you want to bypass the approval step entirely.

Manual Download

Download from Releases.

Building from Source

go build -o smartsh ./cmd/smartsh
go build -o smartshd ./cmd/smartshd

# Cross-platform release archives
./scripts/build.sh        # macOS/Linux
.\scripts\build.ps1       # Windows

Release Flow (Maintainers)

git checkout main && git pull
go test ./...
./scripts/build.sh
git add . && git commit -m "release: vX.Y.Z"
git push origin main
git tag vX.Y.Z
git push origin vX.Y.Z

If auto-release does not trigger:

gh release create vX.Y.Z \
  dist/release/smartsh_darwin_amd64.tar.gz \
  dist/release/smartsh_darwin_arm64.tar.gz \
  dist/release/smartsh_linux_amd64.tar.gz \
  dist/release/smartsh_linux_arm64.tar.gz \
  dist/release/smartsh_windows_amd64.zip \
  dist/release/checksums.txt \
  --title vX.Y.Z --generate-notes

Releases · Issues · License

Built with Go. Made for AI-assisted development.