Manager Universal — Multi-Backend AI Delegation for MCP

A Rust MCP server that delegates coding, reasoning, and toolchain tasks to Claude Code, OpenAI Codex, Google Gemini CLI, or OpenAI GPT API through a single tool surface. One server replaces three separate MCP integrations (claude-runner, codex, gemini-mcp) with unified task lifecycle, persistent state, and a live dashboard.

Manager sits between your orchestration context and disposable coding agents. Your chat session holds goal-level reasoning; manager handles subprocess spawning, output tailing, stall detection, retry escalation, and crash recovery. Long-running delegations survive client restarts. Failed tasks retry with backend escalation. Parallel fan-out collects results without burning orchestration tokens.

Part of CPC (Copy Paste Compute) — a multi-agent AI orchestration platform. Related repos: hands · workflow · local · cpc-paths

Install

Pre-built Binary

1. Download from the latest release:
   • Windows x64 → manager-universal-vX.Y.Z-x64.exe
   • Windows ARM64 → manager-universal-vX.Y.Z-aarch64.exe
2. Rename to manager.exe and place in %LOCALAPPDATA%\CPC\servers\.
3. Add to claude_desktop_config.json:

   {
     "mcpServers": {
       "manager": {
         "command": "%LOCALAPPDATA%\\CPC\\servers\\manager.exe",
         "env": {
           "CPC_DASHBOARD_PORT": "9999",
           "CPC_VOLUMES_PATH": "C:\\My Drive\\Volumes"
         }
       }
     }
   }
   

4. Restart Claude Desktop.

Build from Source

git clone https://github.com/AIWander/manager-universal.git
cd manager-universal
cargo build --release
# binary at target/release/manager.exe

Requires Rust stable. Windows only (toast notifications, PID-liveness checks use Windows APIs).

---

48 Tools

Task Lifecycle (12 tools)

Tool	Description
task_submit	Delegate a prompt to a backend (Claude Code, Codex, Gemini, GPT)
task_status	Get current status, output tail, and metadata for a task
task_output	Stream full stdout/stderr from a running or completed task
task_list	List all tasks with optional status/session filters
task_cancel	Kill a running task's subprocess and mark cancelled
task_poll	Long-poll for task state changes (blocks until update or timeout)
pause_task	Manually pause a running task
resume_task	Resume a paused task (restarts subprocess)
task_retry	Retry a failed task with backend/effort escalation
task_rerun	Re-run a completed task with the same or modified prompt
task_rollback	Rollback a task's file changes using git diff
task_cleanup	Remove completed/failed tasks from state

Session Management (5 tools)

Tool	Description
session_start	Start a persistent interactive session with a backend
session_send	Send a follow-up message to an active session
session_list	List active sessions
session_destroy	Terminate a session and its subprocess
send	Alias for session_send with shorter syntax

Parallel & Routing (4 tools)

Tool	Description
task_run_parallel	Fan out multiple prompts to backends, collect all results
task_route	Auto-route a prompt to the best backend based on task shape
task_decompose	Break a complex prompt into subtasks with dependency graph
task_explain	Explain what a task did, decisions made, files changed

Loaf (Task Groups) (4 tools)

Tool	Description
create_loaf	Create a named task group with phases and acceptance criteria
loaf_update	Update loaf progress, add findings, mark phases done
loaf_status	Get current loaf state including subtask roll-up
loaf_close	Close a loaf with summary and final status

Templates (3 tools)

Tool	Description
template_save	Save a reusable prompt template with variable slots
template_list	List saved templates
template_run	Execute a template with variable substitution

Roles (3 tools)

Tool	Description
role_list	List available agent roles (architect, implementer, tester, etc.)
role_create	Define a custom role with system prompt and constraints
role_delete	Remove a custom role

Extraction & Workflow (4 tools)

Tool	Description
workflow_run	Run a multi-step workflow (sequential task chain)
review_extractions	Review pending extraction candidates from task outputs
extract_workflow	Accept and persist an extraction
dismiss_extraction	Dismiss a pending extraction

Analytics & Monitoring (7 tools)

Tool	Description
status_bar	One-line summary of active tasks, sessions, and system health
task_watch	Watch a task until completion with configurable poll interval
get_analytics	Historical stats: success rates, durations, backend comparison
run_analyzer	Run analysis passes over task history (patterns, failures)
notify	Send a Windows toast notification
configure	Get/set manager runtime configuration
open_terminal	Open a new terminal window at a given path

Dashboard (3 tools)

Tool	Description
dashboard_open	Open the live dashboard in the default browser
dashboard_stop	Stop the dashboard HTTP server
dashboard_status	Get dashboard URL, port, and running state

Backend Shortcuts (3 tools)

Tool	Description
gemini_direct	One-shot Gemini CLI call without full task lifecycle
codex_exec	One-shot Codex call without full task lifecycle
codex_review	Codex code review on a file or diff

---

Dashboard

The embedded dashboard runs at http://127.0.0.1:{port}/ during any active session. Default port is 9218; pin it with CPC_DASHBOARD_PORT. Zones:

• Sessions — active backend sessions with PID, uptime, last activity
• Active Loafs — task groups in progress with phase completion
• Active Operations — running tasks with live output tails
• Last 5 Tools — recent MCP tool calls with timing
• Retry Chains — chevron-linked retry sequences showing escalation path

Find your URL: dashboard_status tool, or read %LOCALAPPDATA%\manager-mcp\dashboard_url.txt.

---

Backends

Backend	Strengths	Typical Use	Speed	Cost
Claude Code	Multi-file edits, tool use, deep reasoning, iterative refinement	Refactors, new features, complex debugging	Medium (30-120s)	High
Codex	One-shot code generation, fast turnaround, good at scripts	Single-file tasks, quick fixes, reviews	Fast (5-20s)	Low
Gemini CLI	Large context window (1M tokens), broad knowledge	Large-context Q&A, doc synthesis, exploration	Medium (15-60s)	Low
GPT API	Structured output, function calling, flexible models	Analysis, classification, structured extraction	Fast (3-15s)	Medium

Backend escalation on retry: Claude Code → Codex → Gemini. Effort escalation: low → medium → high → max.

---

Auto-Pause & Safety

Manager includes a destructive-command heuristic that auto-classifies prompts containing keywords like git push, docker push, terraform apply, cargo build, npm install, delete, --write as potentially destructive.

These tasks default to stall_recovery: manual — if the subprocess dies, the task pauses instead of auto-retrying. A toast notification fires and the task waits for resume_task or task_cancel. Paused tasks aren't failures; check paused_reason and decide.

Override per-task: pass stall_recovery: auto to force auto-retry, or stall_recovery: manual on any task for extra safety.

Retry budget: max_retries defaults to 1. Exhausted budget → failed_max_retries status (task stops permanently). Previous error context is injected into retry prompts.

---

Environment Variables

Variable	Purpose	Default
CPC_DASHBOARD_PORT	Pin dashboard to a specific port	9218
CPC_VOLUMES_PATH	Knowledge-base root path	Auto-detected via cpc-paths
CPC_WORKSPACE_ROOT	Source/workspace root	C:\rust-mcp
OPS_BREADCRUMBS_DIR	Ops server breadcrumb directory	%LOCALAPPDATA%\CPC\ops-data\logs
LOCAL_BREADCRUMBS_DIR	Local server breadcrumb directory	%LOCALAPPDATA%\CPC\local-data\logs
OPENAI_API_KEY	API key for GPT backend	(none)

---

Related Repos

Repo	What it does
hands	Browser, UIA, and vision desktop automation (117 tools)
workflow	API discovery, credential vault, data pipelines, watches (37 tools)
local	Filesystem, shell, git, transforms, sessions
autonomous	Knowledge engine, extractions, Volumes, learning
voice-mcp	Voice input/output for MCP agents

---

Contributing

Issues welcome; PRs considered but this is primarily maintained as part of the CPC stack.

License

Apache 2.0 — see LICENSE.

Copyright 2026 Joseph Wander.

---

Contact

Joseph Wander

• GitHub: github.com/AIWander
• Email: josephwander@gmail.com