Claude Command Runner

Lets Claude actually use your terminal. You ask, Claude types the command into your Warp tab, captures the output, and tells you what happened. It's a Model Context Protocol (MCP) server with 40 tools — command execution, project setups, file watching, SSH, clipboard, environment intelligence — and the same binary works from both Claude Desktop and Warp's built-in AI agent panel. macOS, Swift, open source.

Built for Warp Terminal (free, open source AGPL-3.0). The 5 most powerful tools route commands visibly into your active Warp tab — that's the magic of this MCP. If you don't use Warp yet, grab it here — it pairs with this MCP for what's probably the closest thing to "Claude with a real terminal" you can get right now.

🚀 What's New in v6.0.0 — Warp re-pivot

Triggered by Warp going open source under AGPL-3.0 (github.com/warpdotdev/warp) in May 2026. v6.0 re-establishes Warp as the primary integration target and adds a new install path: Warp's native agent panel.

• Dual-consumer architecture: register claude-command-runner in ~/.warp/.mcp.json and Warp's built-in agent (which uses your configured LLM — Claude Sonnet, Opus, etc.) calls the same tools Claude Desktop calls. Same binary, same code, two consumers. See [docs/WARP_AGENT.md](docs/WARP_AGENT.md).
• Deeplinks replace AppleScript for tab/window operations: open_terminal_tab now uses warp://action/new_tab?path=.... No menu-clicking, no Accessibility permission for the open path.
• OSC 777 emitter (emit_warp_event): build printf invocations that surface structured events into Warp's notification UI. Schema reimplemented from upstream event/v1.rs (no vendoring).
• Workspace profiles → Warp launch configs: save_workspace_profile accepts include_warp_launch_config: true to also write a YAML to ~/.warp/launch_configurations/, making the profile appear in Warp's launch UI.
• Optional shell shim (Tier E, opt-in): helper/install-shim.sh adds zsh/bash hooks that emit preexec/command_finished events to a per-uid Unix socket the MCP listens on. Auto-disables outside Warp panes. Observability surface in v6.0; auto-routing execute_command through it is deferred to v6.0.x.
• Cleanup: ~460 LOC of dead code removed (obsolete Warp DB integration, orphan TCP listener, unreachable background monitor). Tool count corrected: README claimed 30, reality was 36, v6.0 ships 39.
• Reproducible build: swift-sdk pinned to 0.10.x with a deterministic scripts/patch-swift-sdk.sh so a fresh clone builds without manual patches.

Honest caveat: the "Claude" replying in Warp's agent panel is *Warp's* agent (running whichever LLM you configure in Warp's AI settings), not your Claude Desktop instance. Both can call the MCP; they don't share conversations. v6.0 does not bridge them.

What's New in v5.0.0 (history)

v5.0.0: 36 tools, up from 12.

• Clipboard Bridge: Read from and write to the macOS clipboard without leaving the conversation
• macOS Notifications: Get native notifications when long-running commands finish
• Environment Intelligence: Probe your terminal context (git branch, active venv, Node version, Docker containers) in one call
• Output Parsers: Structured JSON parsing for git status, docker ps, test results, and more
• Environment Snapshots: Capture and diff environment variables before/after installs or config changes
• Workspace Profiles: Save and restore project contexts (directory, env vars, common commands) per project
• Multi-Terminal Sessions: Open, name, and send commands to multiple terminal tabs
• Interactive Command Detection: Smart detection of interactive commands (vim, ssh, python REPL) with graceful handling
• File System Watchers: Watch directories for changes and trigger commands automatically
• SSH Remote Execution: Run commands on remote hosts via SSH key authentication

Previous Releases (Included)

• v4.0: Command pipelines, output streaming, reusable templates
• v3.0: Smart auto-retrieve, SQLite history, configurable security

Overview

Claude Command Runner is a dual-consumer MCP server — the same binary serves Claude Desktop and Warp's native agent panel. It enables an LLM (whichever you've configured in either client) to:

• Execute terminal commands directly from conversations
• Chain commands with conditional logic using pipelines
• Stream output in real-time for long builds
• Save and reuse command templates with variables
• Automatically capture output with intelligent timing
• Track command history and patterns
• Read/write the macOS clipboard
• Probe environment context (git, venv, Docker, Node)
• Parse command output into structured JSON
• Manage workspace profiles per project (with optional emission as Warp launch configs)
• Open terminal tabs (via warp:// deeplinks; AppleScript fallback for non-Warp terminals) and send commands to the active tab
• Watch files and trigger commands on changes
• Execute commands on remote hosts via SSH
• (v6.0) Surface tool-execution status to Warp's UI as OSC 777 warp://cli-agent events
• (v6.0, opt-in) Stream clean preexec / command-finished events from your shell to the MCP via a per-uid Unix domain socket

🧭 Which Claude product does this work with?

claude-command-runner is an MCP server, so any Anthropic surface that consumes MCP servers can call its 40 tools — but the value-add varies sharply by consumer, and you should pick deliberately.

Quick decision tree

• You use Warp + Claude Desktop and want commands visible in your terminal? → Register here. Primary use case.
• You use Warp + Claude Code and want everything in one pane? → Don't bother. Claude Code's built-in Bash tool is enough.
• **You use Warp + Claude Code and want commands in a *separate* observable Warp tab?** → Register and use execute_command deliberately.
• You use Warp's native agent panel? → Register in ~/.warp/.mcp.json.
• You don't use Warp? → This isn't the right MCP for you. Look at simpler subprocess-based MCPs.

The "40 tools" headline is honest but slightly misleading: 5 of those tools (the AppleScript-keystroke ones — execute_command, execute_with_auto_retrieve, execute_with_streaming, run_template, send_to_session) are the Warp-routing crown jewels, and 25 of them are pure server-side utilities that any subprocess-based MCP could provide. The remaining 10 sit in between (deeplinks, OSC 777, profiles, file watchers). If the Warp-routing 5 don't appeal, the rest of this MCP is fine but unremarkable.

🎯 Key Features

Command Pipelines

Chain multiple commands with intelligent failure handling:

{
  "steps": [
    {"name": "Build", "command": "swift build", "on_fail": "stop"},
    {"name": "Test", "command": "swift test", "on_fail": "continue"},
    {"name": "Package", "command": "swift build -c release", "on_fail": "stop"}
  ]
}

Failure modes:

• stop – Halt pipeline on failure
• continue – Log error and proceed to next step
• warn – Show warning and continue

Output Streaming

Real-time output for long-running commands:

{
  "command": "swift build -c release",
  "update_interval": 3,
  "max_duration": 180
}

Perfect for:

• Long compilation processes
• Test suites
• Any command that previously "hung" waiting for output

Command Templates

Save reusable patterns with variable substitution:

// Save a template
{
  "name": "swift-release",
  "template": "cd {{project}} && swift build -c release",
  "category": "Swift Development",
  "description": "Build Swift project in release mode"
}

// Run with variables
{
  "name": "swift-release",
  "variables": {"project": "~/GitHub/MyApp"}
}

Templates are stored in ~/.claude-command-runner/templates.json and persist across sessions.

Smart Auto-Retrieve

The execute_with_auto_retrieve command intelligently detects command types and adjusts wait times:

• Quick commands (echo, pwd): 2-6 seconds
• Moderate commands (git, npm): up to 20 seconds
• Build commands (swift build, make): up to 77 seconds
• Test commands: up to 40 seconds

📊 Why Warp Terminal?

Warp Terminal is the primary integration target. The other terminals work for the basics — what Warp uniquely unlocks in v6.0 is documented integration surfaces:

Output capture (/tmp/.json polling) and the 24 tools that don't touch the terminal at all (clipboard, SSH, file watch, env snapshots, etc.) work identically across all terminals.

💡 Get Warp: Download Warp Terminal – it's free, open source (AGPL-3.0), and makes the v6.0-specific surfaces above available to you.

Installation

Prerequisites

• macOS 13.0 or later
• Swift 6.0+ (Xcode 16+)
• Claude Desktop
• A supported terminal (Warp strongly recommended)

Quick Install

1. Clone and build:

git clone https://github.com/M-Pineapple/claude-command-runner.git
cd claude-command-runner

# For the 5 keystroke-routing tools (execute_command, etc.) to work, the build
# must be SIGNED with your code-signing identity. build.sh auto-detects a single
# Apple Development / Developer ID identity; to be explicit (or if you have
# several), export it first — find yours with:
#   security find-identity -v -p codesigning
export CCR_CODESIGN_IDENTITY=""   # optional if auto-detect finds one; persist in ~/.zshrc
./build.sh

Only need the 34 non-keystroke tools (incl. execute_pipeline)? An unsigned build is fine — skip the export.

2. Pick your consumer(s) — you can install for one or both. **v6.0.3 ships a .app bundle wrapper** — point at the binary INSIDE the bundle (the wrapper carries the Info.plist that macOS Sequoia+ needs to prompt for TCC permissions; without it, the 5 keystroke-routing tools silently fail):

A — Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
     "mcpServers": {
       "claude-command-runner": {
         "command": "/path/to/claude-command-runner/.build/release/claude-command-runner.app/Contents/MacOS/claude-command-runner",
         "args": []
       }
     }
   }

B — Warp's native agent (~/.warp/.mcp.json):

{
     "mcpServers": {
       "claude-command-runner": {
         "command": "/path/to/claude-command-runner/.build/release/claude-command-runner.app/Contents/MacOS/claude-command-runner",
         "args": []
       }
     }
   }

See [docs/WARP_AGENT.md](docs/WARP_AGENT.md) for the full guide.

Upgrading from v6.0.0–6.0.2? Edit your existing config file and append .app/Contents/MacOS/claude-command-runner to the path. The legacy bare-binary path still works for the 34 non-keystroke tools, but execute_command / execute_with_auto_retrieve / execute_with_streaming / run_template / send_to_session will silently fail without the bundle path.

3. Grant Accessibility permission (only required for send_to_session keystroke injection in v6.0; tab/window opening uses deeplinks and does not require it):

• Open System Settings → Privacy & Security → Accessibility
• Click + and navigate to claude-command-runner/.build/release/
• Press Cmd+Shift+. to reveal the hidden .build folder
• Select the claude-command-runner binary and toggle it on

Important: macOS tracks permissions by binary identity. After every rebuild (./build.sh), you must remove the old entry and re-add the new binary in Accessibility settings.

4. Restart your client(s) (Claude Desktop and/or Warp).

5. (Optional) Install the shell shim for cleaner block-boundary capture:

helper/install-shim.sh

See helper/shell-shim.zsh / helper/shell-shim.bash for the implementation. Uninstall with helper/uninstall-shim.sh.

Upgrading from a previous version

If you already have a signed install working (TCC permissions granted), upgrading is:

cd claude-command-runner
git pull

# Rebuild with the SAME signing identity you used originally. Without it, build.sh
# falls back to an ad-hoc-signed bundle, macOS sees a new identity, and your
# keystroke (TCC) grants stop applying → error 1002 on execute_command.
# (build.sh auto-detects a single identity; export to be explicit.)
export CCR_CODESIGN_IDENTITY=""   # persist in ~/.zshrc so you don't forget
./build.sh

# Confirm it signed with your cert (NOT adhoc) BEFORE replacing your good bundle:
codesign -dvv .build/release/claude-command-runner.app 2>&1 | grep -E "Authority=Apple|Signature=adhoc"

# Replace the deployed bundle (rm first — cp -R onto an existing .app nests it):
rm -rf "/Applications/Claude Command Runner.app"
cp -R .build/release/claude-command-runner.app "/Applications/Claude Command Runner.app"

Then restart Claude Desktop. Because the bundle keeps the same identifier and is signed with the same certificate, your existing TCC grants (Automation, Input Monitoring, Accessibility, Full Disk Access) carry over — no re-granting needed. If the keystroke tools start failing with error 1002 after an upgrade, you almost certainly rebuilt unsigned; re-sign with your cert and redeploy.

🛡️ macOS Sequoia full setup recipe (the 7 ordered steps)

If execute_command / execute_with_auto_retrieve / execute_with_streaming / run_template / send_to_session fail with osascript is not allowed to send keystrokes (1002) even though you've toggled every panel in System Settings, follow this in order — skipping any step leaves a silent denial somewhere in the chain. The other 34 tools work without any of this; execute_pipeline is a fully-functional substitute if you want to skip the whole TCC saga entirely.

This is the empirically-verified recipe from a real 6-hour debugging session. v6.0.3+ ships the bundle infrastructure that makes this possible; v6.0.4 is this documentation pass.

The denial pattern. macOS Sequoia's TCC and sandbox have layered, *non-obvious* requirements for CLI binaries that drive osascript → System Events → keystroke. The error message is misleading — the actual block usually isn't keystroke permission; it's an earlier preflight check that silently aborts the chain. The five gates, in the order macOS evaluates them:

Step 1 — Have an Apple Development cert (or self-signed Code Signing cert)

If you have a paid Apple Developer account, you already have one (check via security find-identity -v -p codesigning). If not, create a self-signed one:

1. Keychain Access → menu Certificate Assistant → Create a Certificate…

2. Name: claude-command-runner, Identity Type: Self Signed Root, Certificate Type: Code Signing

3. Click Create → Continue through warnings → Done

Export the cert identifier for build.sh to find:

# Get the SHA-1 hash (more reliable than the cert name)
security find-identity -v -p codesigning
# Then in your shell rc (~/.zshrc, ~/.config/fish/config.fish, etc.):
export CCR_CODESIGN_IDENTITY=""

Step 2 — Build (creates the signed .app bundle)

./build.sh

build.sh invokes scripts/make-app-bundle.sh, which wraps the CLI in .build/release/claude-command-runner.app/ with a proper Info.plist (CFBundleIdentifier com.m-pineapple.claude-command-runner, the three required NSXxxUsageDescription strings, LSUIElement=true). If CCR_CODESIGN_IDENTITY is set, the bundle is signed with that cert as a unit — stable cdhash across rebuilds.

Verify:

codesign --display --verbose=4 .build/release/claude-command-runner.app | grep -E 'Identifier|TeamIdentifier|CDHash'
codesign --verify --deep --strict .build/release/claude-command-runner.app  # should succeed silently

Step 3 — Install the bundle into /Applications/ (CRITICAL)

**macOS refuses to prompt for TCC permissions on bundles in .build/release/ or other dev directories.** The bundle must live in /Applications/. Copy it:

cp -R .build/release/claude-command-runner.app "/Applications/Claude Command Runner.app"

Verify the signature survived the copy:

codesign --verify --deep --strict "/Applications/Claude Command Runner.app"

Step 4 — Point your MCP config at the /Applications/ path

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "claude-command-runner": {
      "command": "/Applications/Claude Command Runner.app/Contents/MacOS/claude-command-runner",
      "args": []
    }
  }
}

Mirror in ~/.warp/.mcp.json if you use the Warp Agent path.

Step 5 — Reset stale TCC entries for the bundle ID

If you've been struggling with TCC denials previously, your TCC.db likely has stale Denied entries from earlier rebuilds with different cdhashes. Wipe them for the bundle:

for svc in AppleEvents ListenEvent PostEvent Accessibility; do
    tccutil reset "$svc" com.m-pineapple.claude-command-runner
done

Each line should print "Successfully reset". If you see "no entries", that's also fine — means TCC had nothing recorded yet.

Step 6 — Grant the *three* TCC permissions (Full Disk Access is the surprise one)

In System Settings → Privacy & Security, add /Applications/Claude Command Runner.app to each of:

1. Full Disk Access — *unexpected but mandatory.* macOS sandbox does a preflight check for kTCCServiceSystemPolicyAllFiles before allowing osascript to spawn for keystroke chains. Without FDA on the bundle, the sandbox denies before TCC's AppleEvents check ever fires, and you get the misleading "send keystrokes" error.

2. Input Monitoring — for kTCCServiceListenEvent / kTCCServicePostEvent (synthetic keystroke generation).

3. Accessibility — for the keystroke AppleEvent action itself.

Each grant requires Touch ID / admin password to confirm. Bundle name appears as "Claude Command Runner" in the panels.

Step 7 — Restart Claude Desktop, trigger once, grant the Automation prompt

⌘Q Claude Desktop, reopen. On your first execute_command, macOS may show one more prompt — Automation → "Claude Command Runner wants to control System Events" — click Allow. After that, it's permanent. Future rebuilds don't reset anything (the cert keeps cdhash stable; the bundle keeps the identity stable).

---

Diagnostic: what does TCC see right now?

If something doesn't work after the recipe, the only reliable way to figure out which gate is failing is the TCC log:

log show --predicate 'process == "tccd"' --last 30s --info --debug \
    | grep -E 'AUTHREQ_CTX|AUTHREQ_RESULT|m-pineapple|promptPolicy|Service Policy'

Trigger an execute_command first, then immediately run the above. Read it for:

• service="kTCCServiceXxx" — which permission category is being checked
• promptPolicy = 0 → macOS refuses to even prompt; usually a location problem (bundle not in /Applications/)
• promptPolicy = 2 + Denied (Service Policy) → you're missing the permission for the named service
• AttributionChain: responsible={identifier=com.m-pineapple.claude-command-runner, ...} → good, TCC is identifying us correctly

If you really can't get it working

execute_pipeline is a fully-functional substitute for execute_command:

// instead of execute_command: {"command": "git status"}
// use:
{"steps": [{"command": "git status"}]}

Pure subprocess, no AppleScript, no TCC layer, captures output cleanly, works on any macOS version regardless of permissions. The visible side-effect (command appearing in your Warp tab) is the only thing you lose — for most agent workflows that's not what you actually need anyway.

---

Usage

Available Tools (40)

Core Execution

Clipboard (v5.0)

Notifications (v5.0)

Environment Intelligence (v5.0)

Workspace Profiles (v5.0)

Multi-Terminal Sessions (v5.0)

Interactive Command Detection (v5.0)

File Watching (v5.0)

SSH Remote Execution (v5.0)

Warp v6.0 — deeplinks, OSC 777, shell shim

Example Workflows

Simple Command:

You: "Check my Swift version"
Claude: [execute_with_auto_retrieve: swift --version]
Claude: "You're running Swift 6.0.2"

Build Pipeline:

You: "Build, test, and package my app"
Claude: [execute_pipeline with build → test → package steps]
Claude: "Pipeline complete! Build: ✅ Test: ✅ Package: ✅"

Streaming Long Build:

You: "Build this large project"
Claude: [execute_with_streaming: swift build -c release]
Claude: "Building... [live updates every 3 seconds]"
Claude: "Build completed in 45 seconds!"

Using Templates:

You: "Save a template for deploying to staging"
Claude: [save_template: name="deploy-staging", template="cd {{project}} && ./deploy.sh staging"]

You: "Deploy MyApp to staging"
Claude: [run_template: name="deploy-staging", variables={project: "MyApp"}]

Environment Context (v5.0):

You: "What's my current dev environment?"
Claude: [get_environment_context]
Claude: "You're on branch feature/auth, Python venv active, Node 20.11, 3 Docker containers running."

Workspace Profiles (v5.0):

You: "Save this as my API project profile"
Claude: [save_workspace_profile: name="api-project", directory="~/Projects/api", ...]

You: "Switch to the API project"
Claude: [load_workspace_profile: name="api-project"]

File Watching (v5.0):

You: "Rebuild whenever a Swift file changes"
Claude: [add_file_watch: path="./Sources", pattern="*.swift", command="swift build"]
Claude: "Watching ./Sources for *.swift changes. Will run swift build on each change."

SSH Remote Execution (v5.0):

You: "Check disk space on the staging server"
Claude: [ssh_execute: host="staging.example.com", username="deploy", command="df -h"]
Claude: "Here's the disk usage on staging..."

Configuration

The configuration file is located at ~/.claude-command-runner/config.json:

{
  "terminal": {
    "preferred": "auto",
    "fallbackOrder": ["Warp", "WarpPreview", "iTerm", "Terminal"]
  },
  "security": {
    "blockedCommands": ["rm -rf /", "format"],
    "maxCommandLength": 1000
  },
  "history": {
    "enabled": true,
    "maxEntries": 10000
  },
  "notifications": {
    "enabled": true,
    "soundEnabled": true,
    "showOnSuccess": false,
    "showOnFailure": true,
    "minimumDuration": 10
  },
  "fileWatching": {
    "maxWatchers": 5,
    "defaultDebounce": 2.0,
    "autoExpireMinutes": 60
  },
  "ssh": {
    "defaultTimeout": 30,
    "allowPasswordAuth": false
  },
  "interactiveDetection": {
    "enabled": true,
    "customPatterns": []
  }
}

Templates are stored separately at ~/.claude-command-runner/templates.json.

Workspace profiles are stored at ~/.claude-command-runner/profiles.json.

SSH profiles are stored at ~/.claude-command-runner/ssh_profiles.json.

🤔 Frequently Asked Questions

Q: What's new in v6.0.0?

A: Re-pivot to Warp after Warp went open source. Dual-consumer architecture (register the same binary in ~/.warp/.mcp.json to use it from Warp's native agent panel, in addition to Claude Desktop). warp:// deeplinks replace AppleScript menu-clicking for tab/window operations. New OSC 777 emitter (emit_warp_event) surfaces structured events into Warp's UI. New optional shell shim emits clean preexec/command-finished events to the MCP. Workspace profiles can now also emit Warp-native launch configs. ~460 LOC of dead code removed. Tool count goes from a previously-undercounted 36 (the v5 README claimed 30) up to 39. See CHANGELOG.md and docs/WARP_AGENT.md for the full story.

Q: What was new in v5.0.0?

A: Ten new feature categories bringing the tool count from 12 to 36 (the v5 README under-counted as 30; the dispatch table actually registered 36). Highlights: clipboard integration, macOS notifications, environment intelligence, structured output parsing, workspace profiles, multi-terminal sessions, file watchers, SSH remote execution.

Q: When should I use pipelines vs regular commands?

A: Use pipelines when you need:

• Multiple sequential commands
• Conditional logic (stop on build failure, continue on test failure)
• A summary of all steps with timing
• CI/CD-style workflows

Q: Why does my command "hang" with execute_with_auto_retrieve?

A: For very long commands, use execute_with_streaming instead. It provides real-time output updates and handles commands that run for minutes. This was the main motivation for adding streaming in v4.0.

Q: How do I use templates with multiple variables?

A: Define variables in your template with {{variable_name}} syntax:

{
  "template": "cd {{project}} && git checkout {{branch}} && swift build -c {{config}}"
}

Then provide all variables when running:

{
  "variables": {"project": "~/MyApp", "branch": "main", "config": "release"}
}

Q: Where are my templates stored?

A: In ~/.claude-command-runner/templates.json. They persist across sessions and Claude Desktop restarts.

Q: How long will auto-retrieve wait for my command?

A: It depends on the command type:

• Simple commands: 6 seconds
• Git/npm commands: 20 seconds
• Build commands: 77 seconds
• Unknown commands: 30 seconds

For longer commands, use execute_with_streaming instead.

Q: Can I use this with Terminal.app or iTerm2?

A: Yes, basic command execution works with any terminal. However, automatic output capture and advanced features require Warp Terminal. Get Warp free here.

Q: Is it secure to let Claude run commands?

A: Commands are sent directly to your terminal and execute automatically — there is no manual "press Enter" step. However, you can configure blocked commands and patterns in the config file to prevent dangerous operations. Always review the command Claude proposes before confirming in the chat interface.

Q: What happens if a pipeline step fails?

A: Depends on the on_fail setting:

• stop – Pipeline halts immediately, remaining steps are skipped
• continue – Error is logged, pipeline continues to next step
• warn – Warning is shown, pipeline continues

Q: Can I nest pipelines or run templates inside pipelines?

A: Not directly. You can create templates that contain multiple commands separated by && or ;, or compose by calling run_template and execute_pipeline from the same conversation.

Q: Where is my command history stored?

A: In an SQLite database at ~/.claude-command-runner/claude_commands.db. It tracks all commands, outputs, exit codes, and execution times.

🛠️ Troubleshooting

macOS Permission Error: "osascript is not allowed to send keystrokes" (Error 1002)

This error affects the 5 AppleScript-keystroke-routed tools (execute_command, execute_with_auto_retrieve, execute_with_streaming, run_template, send_to_session). The other 34 tools — including execute_pipeline — are unaffected.

Don't waste hours toggling System Settings panels. This is a layered problem with macOS Sequoia's TCC + sandbox enforcement, and partial fixes leave silent denials in deeper layers. The canonical fix is the **🛡️ macOS Sequoia full setup recipe** above (in the Installation section) — seven ordered steps including: build the .app bundle, install into /Applications/ (not .build/), tccutil reset for the bundle ID, and grant THREE TCC permissions including the surprise one (Full Disk Access on the bundle — sandbox preflights this before allowing the keystroke chain). Follow it top to bottom; skipping any step leaves a hidden denial.

Quick triage — am I hitting this?

Run this immediately after a failed execute_command:

log show --predicate 'process == "tccd"' --last 20s --info --debug | grep -E 'm-pineapple|promptPolicy|Service Policy'

You'll see one of:

• promptPolicy = 0 → bundle isn't in /Applications/ (or no bundle at all). Fix: Steps 2-3 of the recipe.
• promptPolicy = 2 + Denied (Service Policy) for kTCCServiceSystemPolicyAllFiles → missing Full Disk Access on the bundle. Fix: Step 6 of the recipe.
• AttributionChain: responsible={identifier=claude-command-runner, ...} (no m-pineapple) → you're not running from the bundle, you're running the bare binary. Fix: re-run ./build.sh and update your MCP config to the .app path (Steps 2-4 of the recipe).

Workaround if you don't want to bother with the recipe:

// Use execute_pipeline instead of execute_command:
{"steps": [{"command": "git status"}]}

Pure subprocess, no AppleScript, zero TCC. Works on any macOS regardless of permissions. The 34 non-keystroke tools all work too.

Bundle ID Reference: com.m-pineapple.claude-command-runner (this project), com.anthropic.claudefordesktop (Claude Desktop).

---

macOS Accessibility Permission Issues

The MCP binary requires Accessibility permission only for the AppleScript keystroke path used by send_to_session (typing into a specific Warp tab) and the legacy non-Warp paths in execute_command for iTerm/Terminal/Alacritty. Most v6.0 tools work without Accessibility: subprocess-based execution, warp:// deeplinks for tab opening, clipboard, SSH, env snapshots, file watch, profiles, and all 24 server-side tools.

Symptoms (when it does matter):

• Error message: osascript is not allowed assistive access. (-1719)
• send_to_session fails; clipboard, SSH, env-context, deeplink-based open_terminal_tab all work fine

Solution:

1. Open System Settings → Privacy & Security → Accessibility

2. Click the + button and navigate to your claude-command-runner binary:

/path/to/claude-command-runner/.build/release/claude-command-runner

3. The .build folder is hidden by default — press Cmd+Shift+. in Finder to reveal it

4. Toggle the permission on for the binary

Important: macOS tracks Accessibility permissions by binary identity. After every swift build, the binary changes and you must re-add it to the Accessibility list. This only affects keystroke-injection paths — most tools are unaffected.

---

MCP Not Responding

1. Check the client logs (Claude Desktop or Warp). The MCP server runs as a child of the client over stdio — there is no listening TCP port in v6.0.

2. Restart the client (Claude Desktop and/or Warp).

3. Rebuild with ./build.sh. (And re-add the binary to Accessibility if send_to_session is failing.)

Commands Not Appearing in Terminal

1. Ensure Warp/WarpPreview is running

2. Check Claude Desktop logs for errors

3. Verify your MCP configuration path

Streaming Not Updating

1. Check that the command is actually running (not waiting for input)

2. Increase update_interval if updates are too frequent

3. Check /tmp/claude_stream_*.log for output files

Pipeline Steps Skipped Unexpectedly

1. Check the on_fail setting – stop will skip remaining steps

2. Verify each command works individually first

3. Check exit codes in the pipeline summary

Templates Not Saving

1. Ensure ~/.claude-command-runner/ directory exists

2. Check write permissions on templates.json

3. Verify JSON syntax in template definition

Auto-Retrieve Not Working

1. Ensure you're using execute_with_auto_retrieve (not execute_command)

2. Check if command output file exists: ls /tmp/claude_output_*.json

3. For long commands, use execute_with_streaming instead

Database Issues

If commands execute but aren't saved to the database:

1. Check database integrity:

sqlite3 ~/.claude-command-runner/claude_commands.db "PRAGMA integrity_check;"

2. If corrupted, backup and remove:

mv ~/.claude-command-runner/claude_commands.db ~/.claude-command-runner/claude_commands.db.backup
   # Restart Claude Desktop - a new database will be created automatically

Architecture

┌──────────────────┐                     ┌──────────────────┐
  │  Claude Desktop  │                     │   Warp Agent     │
  │   (TS client)    │                     │  (rmcp client)   │
  └────────┬─────────┘                     └────────┬─────────┘
           │ stdio                                  │ stdio
           │ MCP                                    │ MCP
           └──────────────────┬─────────────────────┘
                              ▼
              ┌─────────────────────────────────┐
              │   claude-command-runner v6.0    │
              │      (Swift, MCP server)        │
              │      • 40 tools                 │
              └─────────┬─────────┬─────────────┘
                        │         │
              ┌─────────┘         └──────────────┐
              ▼                                  ▼
   ┌────────────────────┐          ┌─────────────────────────┐
   │   Warp Terminal    │          │   Optional shell shim   │
   │  warp:// deeplinks │◀────────▶│ /tmp/ccr-shell-shim-    │
   │  OSC 777 events    │          │   .sock (NIO)      │
   │  AppleScript (typ) │          │   preexec/finished      │
   └────────────────────┘          └─────────────────────────┘
                        │
                        ▼
               ┌─────────────────┐
               │   Remote Hosts  │
               │    (via SSH)    │
               └─────────────────┘

Two MCP consumers, one server. Tab/window operations use Warp's warp:// URL scheme; status events use OSC 777 (printf-built); typing into a specific tab still uses AppleScript keystrokes (no Warp API exists for that). Optional shell shim adds clean preexec / command_finished events over a per-uid Unix socket.

Contributing

We love contributions! Here's how:

1. Fork the repository

2. Create a feature branch (git checkout -b feature/amazing-feature)

3. Commit your changes (git commit -m 'Add amazing feature')

4. Push to the branch (git push origin feature/amazing-feature)

5. Open a Pull Request

Development Setup

git clone https://github.com/yourusername/claude-command-runner.git
cd claude-command-runner
swift package resolve
swift build

💖 Support This Project

If Claude Command Runner has helped enhance your development workflow or saved you time with intelligent command execution, consider supporting its development:

Your support helps me:

• Maintain and improve Claude Command Runner with new features
• Keep the project open-source and free for everyone
• Dedicate more time to addressing user requests and bug fixes
• Explore new terminal integrations and command intelligence

Thank you for considering supporting my work! 🙏

License

MIT License – see LICENSE file for details

---

Built with ❤️ by 🍍

*If you find this project helpful, give it a ⭐ and try Warp Terminal!*