Agent CLI MCP Server

High-performance multi-agent orchestration server in Rust

One server. Six executors. Full bidirectional control.

---

What It Does

agent-cli-mcp-rust is an MCP server that lets any MCP-capable orchestrator (Claude Code, Cursor, Gemini, etc.) drive multiple AI coding agents through a single unified interface — with built-in directory isolation, secret redaction, tool permission profiles, and session management.

---

Supported Executors

Executor	Binary	Support Level
GitHub Copilot CLI	copilot	Full — prompt, autopilot, fleet, delegate, review, keep-alive
Google Jules	jules	Full — remote sessions, plan approval, API mode, verification
Gemini CLI	gemini	Generic — prompt, interactive, run
OpenAI Codex CLI	codex	Generic — prompt, interactive, run
OpenCode CLI	opencode	Generic — prompt, interactive, run
Anthropic Claude CLI	claude	Generic — prompt, interactive, run

---

Architecture

graph TD
    subgraph Orchestrators["MCP Clients (Orchestrators)"]
        CC[Claude Code]
        GEM[Gemini]
        CUR[Cursor / VS Code]
        COD[Codex]
    end

    subgraph Server["agent-cli-mcp-rust"]
        MCP[MCP Transport\nJSON-RPC over stdio]
        POL[Policy Engine\nDirectory Isolation]
        RED[Redaction Engine\nSecret Scrubbing]
        SES[Session Manager\nAsync I/O]
        CAP[Capability Discovery\nFlag Probing]
    end

    subgraph Executors["AI Coding Agents"]
        COP[Copilot CLI]
        JUL[Jules CLI / API]
        GEMI[Gemini CLI]
        CODEX[Codex CLI]
        OC[OpenCode CLI]
        CLA[Claude CLI]
    end

    Orchestrators -->|stdio| MCP
    MCP --> POL
    MCP --> RED
    MCP --> SES
    MCP --> CAP
    SES -->|stdin/stdout pipes| Executors

Loading

---

Security Model

flowchart LR
    TC[Tool Call] --> DI{Directory\nIsolation}
    DI -->|outside allowed roots| DENY[🚫 Rejected]
    DI -->|within allowed roots| QC{Quarantine\nCheck}
    QC -->|marker present| DENY
    QC -->|clean| EX[Execute]
    EX --> OUT[Raw Output]
    OUT --> RED[Redaction Engine]
    RED -->|scrubs API keys, tokens,\nDB credentials, JWTs| CLEAN[Clean Output]
    CLEAN --> ORCH[Orchestrator]

Loading

Mandatory deny overlay — always blocked regardless of profile:

• memory writes
• vercel deploy --prod
• supabase db reset
• git push --force
• security find-generic-password (macOS Keychain)

---

Tool Permission Profiles

Profile	Use Case
copilot-file-edit	Scoped file edits only
copilot-safe-dev	Dev commands + file edits
copilot-expanded-worktree	Full worktree access
copilot-target-repo	Full repo with build tools

---

Tech Stack

 

---

Installation

Prerequisites

• Rust 1.75+ — install via rustup
• At least one supported executor CLI on your $PATH

Build

git clone https://github.com/pappdavid/agent-cli-mcp-rust.git
cd agent-cli-mcp-rust
cargo build --release
# Binary: target/release/agent-cli-mcp-rust

Quick Verify

echo '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26","capabilities":{},"clientInfo":{"name":"test","version":"0.1.0"}}}' \
  | ./target/release/agent-cli-mcp-rust 2>/dev/null | head -1

---

Configuration

cp config.json ~/.agent-cli-mcp/config.json
export AGENT_CLI_CONFIG=~/.agent-cli-mcp/config.json

Key fields:

Field	Default	Description
allowedRoots	~/Dev,~/projects	Directories the server may operate in
stateDir	~/.agent-cli-mcp	Run logs and session tracking
defaultTimeoutMs	1800000	Executor process timeout (30 min)
maxOutputChars	50000	Max chars returned per tool call

---

Register with MCP Clients

Claude Code / Claude Desktop — add to ~/.claude/settings.json:

{
  "mcpServers": {
    "agent-cli": {
      "command": "/path/to/agent-cli-mcp-rust",
      "env": {
        "AGENT_CLI_ALLOWED_ROOTS": "~/Dev",
        "AGENT_CLI_CONFIG": "~/.agent-cli-mcp/config.json"
      }
    }
  }
}

---

Tool Reference

Discovery & Health

Tool	Description
agent_cli.overview	Compact dashboard of all agent activity
agent_cli.capabilities	Probe CLIs for supported flags and versions
agent_cli.executor_health	Health check across all executors
agent_cli.test_executor_profile	Smoke-test a tool permission profile

Execution

Tool	Description
agent_cli.run	One-shot dispatch (waits for completion)
agent_cli.run_quick	Short-lived run with immediate output
agent_cli.start_run	Background run (returns immediately)
agent_cli.start_session	Long-running interactive session
agent_cli.send_input	Write to session stdin
agent_cli.read_output	Read captured stdout/stderr
agent_cli.list_sessions	List runs and active sessions
agent_cli.kill_session	Terminate a running process

Jules-Specific

Tool	Description
jules.create_session	Create a Jules API session
jules.approve_plan	Approve the latest plan
jules.send_message	Send feedback to a session
jules.request_verification	Ask Jules to verify results
jules.pending_actions	Sessions needing attention

Infrastructure

Tool	Description
agent_cli.create_worktree	Create isolated git worktrees
agent_cli.sanity_check	Scan for mutations or leaks
agent_cli.quarantine	Freeze a directory
agent_cli.collect_artifacts	Collect executor outputs

---

Development

cargo test          # Run tests
cargo fmt           # Format code
cargo clippy        # Lint
RUST_LOG=debug cargo run  # Run with debug logging