The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
npx ecc-install --profile fullFair-code workflow automation platform with native AI capabilities. Combine visual building with custom code, self-host or cloud, 400+ integrations.
npx n8nAn open-source AI agent that brings the power of Gemini directly into your terminal.
npx @google/gemini-cliThese intent pages connect this repository to workflow-first and comparison-first discovery routes.
Supports Claude Code, Codex, Cursor, Mcp
Shows active maintenance signals
Carries strong trust indicators from repository metadata
1076 GitHub stars recorded
| Feature | Description |
|---|---|
| 🤝 Cross-Model/Agent Synergy | Harness strengths from diverse frontier model-powered agents |
| ⚡ Parallel Processing | Multiple agents tackle problems simultaneously |
| 👥 Intelligence Sharing | Agents share and learn from each other's work |
| 🔄 Consensus Building | Natural convergence through collaborative refinement |
| 🖥️ Live Visualization | Interactive Textual TUI with timeline, agent cards, and vote tracking (default). Also available: Web UI, Rich display. |
MassGen configurations are organized by features and use cases. See the Configuration Guide for detailed organization and examples.
Quick navigation:
See MCP server setup guides: Discord MCP | Twitter MCP
For detailed configuration of all supported backends (OpenAI, Claude, Gemini, Grok, etc.), see:
MassGen supports an interactive mode where you can have ongoing conversations with the system:
# Start interactive mode with a single agent (no tool enabled by default)
uv run python -m massgen.cli --model gpt-5-mini
# Start interactive mode with configuration file
uv run python -m massgen.cli \
--config massgen/configs/basic/multi/three_agents_default.yaml
Interactive Mode Features:
""" or ''' to enter multi-line messages/help or /h - Show available commands/status - Display current system status/config - Open the configuration file/clear or /reset - Clear conversation history and start fresh/quit, /exit, or /q - Exit the session (or press Ctrl+C)Watch the recorded demo:
The system provides multiple ways to view and analyze results:
Watch an example here:
All sessions are automatically logged with detailed information for debugging and analysis.
Real-time Interaction:
r during execution to view the coordination table in your terminal.massgen/
└── massgen_logs/
└── log_YYYYMMDD_HHMMSS/ # Timestamped log directory
├── agent_<id>/ # Agent-specific coordination logs
│ └── YYYYMMDD_HHMMSS_NNNNNN/ # Timestamped coordination steps
│ ├── answer.txt # Agent's answer at this step
│ ├── context.txt # Context available to agent
│ └── workspace/ # Agent workspace (if filesystem tools used)
├── agent_outputs/ # Consolidated output files
│ ├── agent_<id>.txt # Complete output from each agent
│ ├── final_presentation_agent_<id>.txt # Winning agent's final answer
│ ├── final_presentation_agent_<id>_latest.txt # Symlink to latest
│ └── system_status.txt # System status and metadata
├── final/ # Final presentation phase
│ └── agent_<id>/ # Winning agent's final work
│ ├── answer.txt # Final answer
│ └── context.txt # Final context
├── coordination_events.json # Structured coordination events
├── coordination_table.txt # Human-readable coordination table
├── vote.json # Final vote tallies and consensus data
├── massgen.log # Complete debug log (or massgen_debug.log in debug mode)
├── snapshot_mappings.json # Workspace snapshot metadata
└── execution_metadata.yaml # Query, config, and execution details
coordination_table.txt): Complete visualization of multi-agent coordination with event timeline, voting patterns, and consensus buildingcoordination_events.json): Structured JSON log of all events (started_streaming, new_answer, vote, restart, final_answer)vote.json): Final vote tallies, winning agent, and consensus informationexecution_metadata.yaml): Original query, timestamp, configuration, and execution context for reproducibilityagent_outputs/): Complete output history and final presentations from all agentsmassgen.log): Complete system operations, API calls, tool usage, and error traces (use --debug for verbose logging)→ For comprehensive logging guide and debugging techniques, see Logging & Debugging
ApprovalProviderrisk-baseddeny-allallow-allrole presets (read-only/researcher deny writes+shell, also empties the agent's SRT writable set), an append-only JSONL ApprovalLedger, a runaway-loop ApprovalBudget, and always-grant persistence to settings.local.jsonPermissionGuardrailSection, injected only when the engine is active): follow the guardrails, don't circumvent a denial, surface-and-ask — while keeping ask a sanctioned path. Authority is established by channel (only the system prompt is authoritative). Denied tool calls now render as first-class failed tool events (with the command) in the TUI/WebUI timelinepermissions: block is 100% unchanged; native backends (claude_code/codex) report INACTIVE rather than silently inert. All under TDD; live-verified that the prompt is best-effort alignment (a model evaded the regex egress classifier via \c\u\r\l / python urllib), so the OS sandbox remains the load-bearing enforcement✅ OS-Level Agent Sandboxing (v0.1.96): Added a real OS-level execution sandbox (command_line_execution_mode: srt) wrapping agent command/code execution in Anthropic's sandbox-runtime (bubblewrap/Seatbelt), with OS-enforced filesystem + network isolation derived from the same PathPermissionManager policy as the app layer (defense in depth), configurable read confinement (confined/strict/open, deny-all network), a hardened key-agnostic permission hook, native-sandbox-backend srt→local degrade, and subagent inheritance.
✅ Steering Improvements (v0.1.95): Extended mid-stream steering from a UI-only capability into a programmatic, headless one — send_steering_message() drops guidance into a file inbox (--inbox-dir) routed through the same set_pending_input chokepoint the TUI/WebUI use — and upgraded Codex/Antigravity to interrupt-and-resume the in-flight turn (codex exec resume / agy --continue) instead of waiting for a round boundary, with expires_at-guarded MCP-hook payload IPC and the Antigravity --model flag wired through.
✅ Parallelism Hardening — Engineering Health (v0.1.94): Moved the peer-context snapshot copy off the event loop (worker thread via asyncio.to_thread) backed by immutable, versioned snapshots (SnapshotVersionStore) with atomically-repointed symlinks and refcounted readers, eliminating the read-during-write race; fixed lost peer-answer revisions (R1), lost background-subagent results (R2/R3), leaked trace tasks (R4), and cancel-without-await teardown (R5); surfaced worktree-isolation degradation (D2); and unified the mid-stream injection paths (A1). No per-backend functionality changes.
✅ CLI Package Decomposition & Pydantic Config Migration (v0.1.93): Split the monolithic 12k-line cli.py into an 18-module massgen/cli/ package, migrated config classes to pydantic.dataclasses with Literal-typed modes the validator derives from, consolidated provider-exclusion lists, removed ~8.7k lines of dead legacy code from the wheel, and hardened the test-signal/type-checking tooling. Internal-quality release, no runtime behavior changes.
✅ Orchestrator Collaborator Refactor & Parallel Search MCP (v0.1.92): Reduced orchestrator.py from 21,599 to 8,574 lines by extracting 49 lazy collaborators, split Textual display helpers into sibling modules, added characterization coverage, and introduced a Parallel Web Search MCP registry entry plus example config.
✅ Config Reliability & Hook Safety (v0.1.91): Centralized coordination, timeout, and orchestrator runtime parsing; strict unknown-key validation; checklist runtime control wiring; and safer Gemini/Codex native hook path permission precedence.
✅ Discriminative Criteria Refinements & Checklist Calibration (v0.1.90): Improved checklist-gated refinement quality with discriminative-power pruning, per-criterion feedback, position-bias counterbalancing, deterministic tie-breaking, a unified checklist gate, shared score parsing utilities, and fast-iteration config updates.
✅ Antigravity CLI Full Integration & Hardening (v0.1.89): Completed the follow-up Antigravity integration pass with workflow-mode parity, auth checks, workspace project anchoring, standalone hooks.json, and prompt affordance gating.
✅ Antigravity CLI Backend (v0.1.88): Added the first antigravity_cli backend wrapping Google's agy binary, with workspace-local .antigravity/ config isolation, MCP config translation, native-hook adapter support, and runnable single-agent / mixed Gemini + Antigravity examples.
✅ Framework Comparisons & llms.txt (v0.1.87): Added CrewAI, LangGraph, and AutoGen/AG2 comparison pages, plus llms.txt and generated llms-full.txt for AI-agent documentation discovery. Also fixed bootstrap_subagent single-shot behavior with refine=False.
✅ bootstrap_subagent Discriminator + Codex MCP Approval Fix (v0.1.86): Variant B is now functional — the orchestrator runs an in-process LLM critic between rounds, merges critic-proposed criteria into the accumulator, and augments the next round's checklist. Codex MCP tool calls under codex exec now write both approval bypasses needed for non-interactive runs.
✅ Discriminative Criteria Emergence (criteria_mode) (v0.1.85): New orchestrator.coordination.criteria_mode lets evaluation criteria emerge from observed gaps across rounds. bootstrap_inline is fully functional on all backends with checklist tool support, with proposed_criteria persisted, deduped, capped, and merged into the next round's effective checklist.
✅ TUI Consensus Map (v0.1.84): Compact visual map below the agent status ribbon during multi-agent runs that summarizes coordination state — agent nodes with latest answer labels, vote arrows, current leader, winner state — driven by existing coordination events without backend schema changes.
✅ In-Session Standalone Checkpoint MCP Integration (v0.1.83): The standalone checkpoint MCP server can now run inside a normal MassGen single-agent session via coordination.standalone_checkpoint config block. Enhanced TUI tool card visualization distinguishes primary checkpoint operations from system tasks.
✅ TUI Copy Mode & Checkpoint Quality Improvements (v0.1.82): Ctrl+Shift+S copy mode toggle for native text selection. Checkpoint quality criteria with selective branch depth scoring, optional workspace context for reviewer agents, and single-checkpoint agent recovery guidance.
✅ Multi-Region Circuit Breaker Failover (v0.1.81): LLM circuit breaker fails over to backup regions when the primary trips OPEN, with automatic recovery when the primary returns to healthy. Completes the circuit breaker series (Phase 1-6).
✅ Adaptive Circuit Breaker & Checkpoint Modes (v0.1.80): Circuit breaker Phase 5 with self-tuning adaptive thresholds. New standalone checkpoint modes: single checkpoint and draft plan verify.
✅ Fast Mode Speed Control & Broader Checkpoint Framing (v0.1.79): New fast mode options for fine-grained speed vs. quality control. Checkpoint framing broadened from safety-only to high-stakes and coordinated phases.
✅ Circuit Breaker Distributed Store (v0.1.78): Pluggable distributed state store for the LLM circuit breaker — share state across workers/processes. In-memory (zero-deps) and Redis-backed implementations with atomic state transitions.
✅ Answer Now Button (v0.1.77): New "Answer Now" button lets agents submit answers more quickly, both within a round, and bypassing additional refinement rounds when quality is sufficient.
✅ Exa Search & Circuit Breaker Observability (v0.1.76): New Exa AI-powered search tool for MCP. Circuit breaker Phase 3 with observability. Checkpoint agent instructions and Docker dependency fixes.
✅ Codex Hooks & Checkpoint WebUI (v0.1.75): Hybrid hook system for Codex backend. Checkpoint workflows auto-launch WebUI for visual monitoring. Standalone checkpoint MCP server docs with safety policy integration.
✅ Checkpoint Improvements & Tool Call Fixes (v0.1.74): Major improvements to standalone checkpoint MCP server. Fix for duplicate tool calls in ChatCompletions and Response API backends.
✅ Eval Criteria Evolver & Checkpoint Objectives (v0.1.73): New eval criteria evolver subagent that evolves criteria across rounds. Initial draft of checkpoint objective mode for safety planning of irreversible actions.
✅ Grok Backend Update & Circuit Breaker Phase 2 (v0.1.72): Grok backend update with latest improvements. LLM API circuit breaker extended to ChatCompletions, Response API, and Gemini backends (was Claude-only).
✅ Trace Memory & Evaluation Polish (v0.1.71): Trace analyzer subagents launch in background after each round to write insights from execution traces into memory. Improved evaluation criteria generation and system prompt tuning.
✅ Evaluation Criteria Redesign (v0.1.70): Redesigned three-tier evaluation criteria with anti-pattern definitions and aspiration statements. Improved checklist-gated evaluation. Fast iteration mode, WebUI review modal, and background trace analysis.
✅ WebUI Automation & Improved Skill (v0.1.69): WebUI automation auto-starts without browser interaction. MassGen skill redesign for increased usability and WebUI integration. Quickstart Wizard rework and Workspace Browser expansion.
✅ Checkpoint Mode (v0.1.68): New checkpoint coordination mode with delegator pattern — main agent plans solo then delegates to team via checkpoint() tool. LLM API circuit breaker for 429 handling. WebUI checkpoint support. LiteLLM supply chain fix.
✅ Modernized WebUI (v0.1.67): Complete WebUI redesign with inline final answers, keyboard shortcuts, and Zustand state management. RoundBudgetGuardHook for per-round cost control. Unified parallel pre-collab phases. Regression guard.
✅ Step Mode (v0.1.66): New --step CLI mode for external orchestrators. Powers massgen-refinery plugin step mode. Codex Windows UTF-8 fixes and console text sanitization.
✅ MassGen Refinery Plugin (v0.1.65): Standalone MCP servers (quality, workflow, media) bring MassGen's checklist-based evaluation to Claude Code through the massgen-refinery plugin. Single-agent refinement working; multi-agent experimental.
✅ Gemini CLI Backend (v0.1.64): Gemini CLI as a first-class backend with session persistence, MCP tools, and Docker support. WebSocket streaming for OpenAI Response API. Execution trace analyzer subagent. Copilot Docker mode.
✅ Ensemble & Contracts (v0.1.63): Subagent ensemble pattern with disable_injection and defer_voting_until_all_answered as defaults. Round evaluator transformation pressure and success contracts. Lighter refinement for subagents. Killed agent handling.
✅ MassGen Skill & Viewer (v0.1.62): General-purpose multi-agent skill with 4 modes (general, evaluate, plan, spec) for Claude Code and other AI agents. Session viewer for real-time observation. Backend improvements for Claude Code, Codex, and Copilot. Headless and web quickstart modes.
✅ Round Evaluator Paradigm (v0.1.61): New round evaluator subagent type that automatically spawns evaluator subagents after each new answer to provide detailed feedback as input to the next round. Major orchestrator refactoring with improved evaluation prompts, task plan injection, and subagent fixes.
✅ Multimodal Tools, Subagent Enhancements & GPT-5.4 (v0.1.60): Rewritten read_media with clearer schema and MediaCallLedgerHook. Subagent enhancements with inherit_spawning_agent_backend, final_answer_strategy, per-agent subagent_agents. GPT-5.4 as default OpenAI flagship. Decomp mode cooperates with checklist workflow. Codex prompt caching fix.
✅ Quality Round Improvements (v0.1.59): Auto-add improvements to task plan, plan review enhancements. Better eval gen config, checklist fixes, Gemini tool name normalization for MCP. Subagent behavior adjustments, Docker skill write access fixes. Video gen skill adjustments and impact metric restoration.
✅ Comprehensive Multimodal Revamp (v0.1.58): ElevenLabs TTS/STT, Nano Banana 2 image generation, Grok multimedia generation, media generation skills, and multi-turn image editing. Nvidia NIM backend. Quality rethinking subagent. Smarter checklists with improve/preserve listings. CLI mode flags and logging architecture refactor.
✅ Delegated Subagent Protocol & Builder Subagent (v0.1.57): File-based delegation protocol for container-to-host subagent spawning. New builder subagent type for large artifact generation with fresh context. Substantiveness tracking for smarter convergence. Claude Code reasoning parameters for updated SDK.
✅ Spec Plan Mode & Targeted Messaging (v0.1.56): Formal requirements specification with plan_mode="spec" and TUI spec mode support. Targeted agent-to-agent messaging via target_agents parameter. Critic subagent for quality assessment. Media conversation continuity for follow-up image analysis. Codex OAuth login fix.
✅ Specialized Subagent Types & Dynamic Evaluation Criteria (v0.1.55): Discovery-based subagent roles (evaluator, explorer, researcher, novelty) via SUBAGENT.md frontmatter. GEPA-inspired task-specific evaluation criteria with core/stretch gates. Native backend image routing. Configurable video frame extraction.
✅ Subagent Messaging & Copilot SDK Backend (v0.1.54): Runtime messaging to steer running background subagents. New GitHub Copilot backend via copilot-sdk with native MCP support. Gemini 3.1 Pro support. Per-agent injection targeting.
✅ Background Tool Execution (v0.1.53): Non-blocking lifecycle tools for long-running work (start, monitor, wait, cancel, list). Planning task verification requirements. TUI background job indicators and lifecycle controls. Subagent infrastructure groundwork with Evaluator and Explorer types.
✅ Final Answer Modal & Coordination Quality Gates (v0.1.52): Dedicated final answer modal with tabbed answer and workspace/review interface. Substantive gate prevents low-value iteration rounds. Novelty injection combats premature convergence. Agent identity versioning for answer provenance tracking.
✅ Reviewing Coordination & Change Documents (v0.1.51): Review modal with multi-file diff visualization. Decision journal system for multi-agent coordination traceability. Changedoc-anchored evaluation checklists with gap reports. Drift conflict policy for safer change application. --cwd-context CLI flag.
✅ Chunked Plan Execution & Skill Lifecycle Management (v0.1.50): Chunked plan execution for safer long-form task completion with progress checkpoints. Skill lifecycle management with consolidation, organizer, and previous-session skill loading. Iterative planning review modal. Responsive TUI mode bar. Worktree improvements with branch accumulation and cross-agent diff visibility.
✅ Coordination Quality: Log Analysis TUI, Fairness Gate & Checklist Voting (v0.1.49): Log analysis mode built into TUI mode bar for in-app run analysis. Fairness gate prevents fast agents from dominating coordination. Checklist voting tool for structured quality evaluation. Automated testing infrastructure with CI/CD and SVG snapshot baselines.
✅ Decomposition Mode & Worktree Isolation (v0.1.48): New decomposition coordination mode that decomposes tasks into subtasks assigned to individual agents with a presenter role, git worktree-based isolation for agent file writes with review modal, quickstart wizard Docker setup with animated pull progress, stop tool for agent completion signaling
✅ Codex Backend & TUI Theme Refactoring (v0.1.47): New Codex backend for OpenAI Codex CLI with local and Docker execution, NativeToolMixin for shared tool handling, TUI theme system refactored to palette-based architecture with dark and light variants, per-agent voting sensitivity configuration
✅ Subagent TUI Streaming & Event Architecture Refactor (v0.1.46): Interactive preview cards that expand to full timeline views with real-time event streaming, unified event pipeline with single source of truth for display creation, enhanced final presentation with workspace visualization and winning agent highlighting, fixed banner display and tool call ID handling
✅ TUI as Default & Config Migration (v0.1.45): Textual Terminal UI now launches by default with automatic rich_terminal to textual_terminal migration, setup wizard generates TUI configs, legacy Rich display accessible via --display rich flag
✅ Execute Mode for Independent Plan Selection (v0.1.44): Mode cycling through Normal → Planning → Execute via Shift+Tab or mode bar, plan selector browsing up to 10 recent plans with timestamps, view full plan modal with complete task breakdown, empty submission for plan execution, context path preservation between planning and execution phases, enhanced case studies with interactive setup guides, TUI performance optimizations with viewport-based rendering
✅ Tool Call Batching & Interactive Case Studies (v0.1.43): Consecutive MCP tool calls grouped into collapsible tree views with "+N more" indicators and click-to-expand. New interactive case studies page with side-by-side SVG comparisons. PlanOptionsPopover for browsing plans and selecting depth. Quoted path support for paths with spaces. Final presentation display and TUI polish fixes.
✅ TUI Visual Redesign & Human Input Queue (v0.1.42): Modern "Conversational AI" aesthetic with rounded corners, redesigned agent tabs with dot indicators, adaptive tool cards, polished modals. New HumanInputHook for injecting messages to agents mid-stream with thread-safe per-agent tracking. AG2 single-agent coordination fix.
✅ Async Subagent Execution (v0.1.41): Background subagent execution with async_=True for non-blocking parallel work, poll for completion and retrieve results, per-round timeout control with subagent_round_timeouts config, extended subagent parameters for timeout and concurrency control
✅ Textual TUI Interactive Mode (v0.1.40): Interactive terminal UI with --display textual for real-time agent streaming, comprehensive modals for costs/votes/workspace/answers, context path injection with @path/to/file syntax, human feedback integration via prompt modals
✅ Plan and Execute Workflow (v0.1.39): Complete plan-then-execute workflow with --plan-and-execute for autonomous planning and execution, --execute-plan to run existing plans without re-planning, task verification workflow with verified status and verification groups for batch validation, plan storage system in .massgen/plans/ with frozen snapshots and execution tracking, Response API function call message sanitization fixes
✅ Task Planning & Two-Tier Workspaces (v0.1.38): Task planning mode with --plan flag for structured work breakdown (plan-only, no auto-execution), git-backed two-tier workspaces separating scratch exploration from final deliverables, automatic CLAUDE.md/AGENTS.md discovery for project context, batch image analysis with multi-image comparison, circuit breaker for timeout denial loops, Docker health monitoring
✅ Execution Traces & Thinking Mode (v0.1.37): Full execution history preserved as execution_trace.md for compression recovery and cross-agent coordination, Claude Code and Gemini reasoning content streaming buffer integration, standardized agent labeling across all backends
✅ @path Context Handling & Hook Framework (v0.1.36): Inline file picker with @path syntax and autocomplete, PreToolUse/PostToolUse hooks for permission validation and content injection, global and per-agent hook registration, built-in MidStreamInjectionHook and HighPriorityTaskReminderHook, Claude Code hooks compatibility, improved Docker resource management
✅ Log Analysis CLI & Logfire Observability (v0.1.35): massgen logs analyze command with prompt mode and multi-agent self-analysis, Logfire workflow attributes for round context and vote reasoning, direct_mcp_servers config for keeping specific MCPs as protocol tools, improved tool handling for unknown tools and vote-only mode fixes
✅ OpenAI-Compatible Server & Model Discovery (v0.1.34): Local HTTP server with massgen serve compatible with any OpenAI SDK client, dynamic model discovery for Groq and Together backends via authenticated API calls, WebUI file diffs and answer refresh polling, subagent status tracking and cancellation recovery improvements
✅ Reactive Context Compression & Streaming Buffers (v0.1.33): Automatic conversation compression when context length errors occur, streaming buffer system tracking partial responses for recovery, file overwrite protection in write_file tool, task plan duplicate prevention, Grok MCP tools visibility fix, Gemini vote-only mode fix, GPT-5 model behavior improvements
✅ Multi-Turn Session Export & Per-Attempt Logging (v0.1.32): Turn range selection for session export (--turns), workspace export controls (--no-workspace, --workspace-limit), Logfire moved to optional [observability] extra, per-attempt isolated log files with handler reconfiguration, automatic DOCX/PPTX/XLSX to PDF conversion for session sharing
✅ Logfire Observability & Azure Tool Streaming (v0.1.31): Optional Logfire integration with automatic LLM instrumentation for OpenAI, Claude, and Gemini backends, Azure OpenAI tool calls yielded as structured chunks, --logfire CLI flag and MASSGEN_LOGFIRE_ENABLED environment variable
✅ OpenRouter Web Search & Persona Diversity (v0.1.30): Native web search via OpenRouter plugins with enable_web_search, persona diversity modes (perspective/implementation) with phase-based adaptation, Azure multi-endpoint auto-detection, environment variable expansion with ${VAR} syntax
✅ Subagent System & Tool Metrics (v0.1.29): Spawn parallel child MassGen processes with isolated workspaces and automatic result aggregation, enhanced tool metrics with per-call averages and min/max/median distribution, CLI per-agent system messages via massgen --quickstart
✅ Unified Multimodal Tools & Artifact Previews (v0.1.28): Consolidated read_media tool for image/audio/video analysis, unified generate_media tool for media creation (images, videos, audio), Web UI artifact previewer for PDFs/DOCX/PPTX/images/HTML/SVG/Markdown/Mermaid, OpenRouter tool-capable model filtering, Azure OpenAI fixes
✅ Session Sharing & Log Analysis (v0.1.27): Session sharing via GitHub Gist with massgen export, log analysis CLI with massgen logs command, per-LLM call timing metrics, Gemini 3 Flash model support, enhanced CLI config builder with per-agent web search and system messages
✅ Web UI Setup & Shadow Agent Depth (v0.1.26): Docker diagnostics module, Web UI setup wizard with guided first-run experience, shadow agent response depth for test-time compute scaling, GPT-5.1-Codex family models
✅ UI-TARS & Evolving Skills (v0.1.25): ByteDance's UI-TARS-1.5-7B for GUI automation, GPT-5.2 model support, evolving skill creator system with session persistence, enhanced Textual terminal with adaptive layouts
✅ Multi-Backend Cost Tracking (v0.1.24): Real-time token counting for OpenRouter, xAI/Grok, Gemini, and Claude Code backends with /inspect c cost breakdown showing per-agent token usage, aggregated session cost totals with improved display formatting
✅ Turn History Inspection & Web UI Automation (v0.1.23): Interactive /inspect commands for reviewing turn details with menu navigation, AutomationView component for programmatic monitoring, SessionMountManager for Docker container persistence across turns, flag-based cancellation with terminal restoration, run_async_safely() for nested event loop handling
✅ Shadow Agent Architecture (v0.1.22): Lightweight shadow agents respond to broadcasts in parallel without interrupting parent work, inheriting full conversation history and current turn context via asyncio.gather() parallelization
✅ Graceful Cancellation & Session Resumption (v0.1.21): Ctrl+C saves partial progress during coordination, cancelled sessions resume with --continue preserving agent answers and workspaces
✅ Web UI & Auto Docker Setup (v0.1.20): Browser-based real-time visualization with React frontend, WebSocket streaming, timeline views, and workspace browsing. Automatic Docker container setup for computer use agents with pre-configured X11 virtual display, xdotool, Firefox, Chromium, and scrot
✅ LiteLLM Integration & Claude Strict Tool Use (v0.1.19): MassGen as LiteLLM custom provider with run() and build_config() programmatic API, Claude strict tool use with structured outputs, Gemini exponential backoff for rate limit resilience
✅ Agent Communication System (v0.1.18): Human broadcast Q&A via ask_others() tool with three modes, blocking execution with inline response delivery, session-persistent Q&A history
✅ Claude Advanced Tooling (v0.1.18): Programmatic tool calling via enable_programmatic_flow flag, server-side tool discovery via enable_tool_search with regex or bm25 variants
✅ Textual Terminal Display (v0.1.17): Interactive terminal UI using the Textual library with dark/light themes, multi-panel layout for agents and orchestrator, real-time streaming with syntax highlighting, content filtering for critical patterns
✅ Terminal Evaluation & Cost Tracking (v0.1.16): Automated VHS recording with AI-powered terminal display evaluation, LiteLLM integration for accurate pricing across 500+ models with reasoning/cached tokens support, memory archiving for multi-turn session persistence, four self-evolution skills for MassGen development
✅ Persona Generation & Docker Distribution (v0.1.15): Automatic persona generation for agent diversity with multiple strategies (complementary, diverse, specialized, adversarial), GitHub Container Registry integration with ARM support, custom tools in isolated Docker containers for security, MassGen pre-installed in Docker images
✅ Parallel Tool Execution & Gemini 3 Pro (v0.1.14): Configurable concurrent tool execution across all backends with asyncio-based scheduling, Gemini 3 Pro integration with function calling, interactive quickstart workflow, MCP registry client for server metadata
✅ Code-Based Tools & MCP Registry (v0.1.13): CodeAct paradigm implementation with tool integration via importable Python code reducing token usage by 98%, MCP server registry with auto-discovery and on-demand loading, TOOL.md documentation standard
✅ NLIP Integration & Skills System (v0.1.13): Advanced tool routing with Natural Language Interface Protocol across Claude, Gemini, and OpenAI backends, cross-platform automated skills installer for openskills CLI, Anthropic skills, and Crawl4AI
✅ System Prompt Architecture Refactoring (v0.1.12): Hierarchical system prompt structure with XML-based formatting for Claude, improved LLM attention management
✅ Semtools & Serena Skills (v0.1.12): Semantic search via embedding-based similarity, symbol-level code understanding via LSP integration, local execution mode for non-Docker environments
✅ Multi-Agent Computer Use (v0.1.12): Enhanced Gemini computer use with Docker integration, VNC visualization, multi-agent coordination combining Claude (Docker/Linux) and Gemini (Browser)
✅ Skills System (v0.1.11): Modular prompting framework with SkillsManager for dynamic skill loading, automatic discovery with always/optional categories, file search skill, Docker-compatible mounting
✅ Memory MCP Tool & Filesystem Integration (v0.1.11): MCP server for memory management with markdown-based storage, short-term/long-term memory tiers, automatic workspace persistence, orchestrator integration for cross-agent memory sharing, enhanced Windows support for long system prompts
✅ Rate Limiting System (v0.1.11): Multi-dimensional limiting (RPM, TPM, RPD) for Gemini models with configurable thresholds, YAML-based configuration, CLI integration with --enable-rate-limiting flag, asyncio lock fix for event loop reuse
✅ Framework Interoperability Streaming (v0.1.10): Real-time intermediate step streaming for LangGraph and SmoLAgent with log/output distinction, enhanced debugging for external framework reasoning steps
✅ Docker Configuration Enhancements (v0.1.10): Nested authentication with separate mount and environment variable arrays, custom image support via Dockerfile.custom-example, automatic package installation
✅ Universal Workspace Isolation (v0.1.10): Instance ID generation extended to all execution modes ensuring safe parallel execution, enhanced workspace path uniqueness across concurrent sessions
✅ Session Management System (v0.1.9): Complete session state tracking and restoration with SessionState dataclass and SessionRegistry for multi-turn persistence across CLI invocations, workspace continuity preserving agent states and coordination history between turns
✅ Computer Use Tools (v0.1.9): Native Claude and Gemini computer use API integration for browser and desktop automation with screenshot analysis and action generation, lightweight browser automation for specific tasks without full computer use overhead
✅ Fuzzy Model Matching (v0.1.9): Intelligent model name search with approximate inputs (e.g., "sonnet" → "claude-sonnet-4-5-20250929"), model catalog system with curated lists across providers, enhanced config builder with automatic model search
✅ Backend Capabilities Expansion (v0.1.9): Comprehensive backend registry with detailed specifications for all providers, audio/video support, hardware acceleration, unified access across diverse model families, enhanced memory update logic focusing on actionable patterns
✅ Automation Mode for LLM Agents (v0.1.8): Complete infrastructure for running MassGen inside LLM agents with SilentDisplay class for minimal output (~10 lines vs 250-3,000+), real-time status.json monitoring updated every 2 seconds, meaningful exit codes (0=success, 1=config error, 2=execution error, 3=timeout, 4=interrupted), automatic workspace isolation for parallel execution, meta-coordination capabilities allowing MassGen to run MassGen
✅ DSPy Question Paraphrasing Integration (v0.1.8): Intelligent question diversity for multi-agent coordination with semantic-preserving paraphrasing module supporting three strategies (diverse/balanced/conservative), automatic semantic validation to ensure meaning preservation, thread-safe caching system with SHA-256 hashing, support for all backends as paraphrasing engines, orchestrator integration for automatic question variant distribution
✅ Agent Task Planning System (v0.1.7): MCP-based planning server with task lifecycle management, dependency tracking with automatic validation and blocking, status transitions between pending/in_progress/completed/blocked states, orchestrator integration for plan-aware multi-agent coordination
✅ Background Shell Execution (v0.1.7): Persistent shell sessions for long-running commands with BackgroundShell class supporting async execution, real-time output streaming and monitoring, automatic timeout handling, enhanced code execution server with background capabilities
✅ Preemption Coordination (v0.1.7): Agents can interrupt ongoing coordination to submit better answers without full restart, partial progress preservation during preemption, enhanced coordination tracker logging preemption events
✅ Framework Interoperability (v0.1.6): AG2 nested chat, LangGraph workflows, AgentScope agents, OpenAI Assistants, and SmoLAgent integrated as custom tools with cross-framework collaboration and streaming support for AG2
✅ Configuration Validator (v0.1.6): Comprehensive YAML validation with ConfigValidator class, pre-commit integration, and detailed error messages with actionable suggestions
✅ Unified Tool Execution (v0.1.6): ToolExecutionConfig dataclass standardizing tool handling across ResponseBackend, ChatCompletionsBackend, and ClaudeBackend with consistent error reporting
✅ Gemini Backend Simplification (v0.1.6): Removed gemini_mcp_manager and gemini_trackers modules, consolidated code reducing codebase by 1,598 lines
✅ Memory System (v0.1.5): Long-term semantic memory via mem0 integration with fact extraction and retrieval across sessions, short-term conversational memory for active context, automatic context compression when approaching token limits, cross-agent memory sharing with turn-aware filtering, session management for memory isolation and continuation, Qdrant vector database integration for semantic search
✅ Multimodal Generation Tools (v0.1.4): Create images from text via DALL-E API, generate videos from descriptions, text-to-speech with audio transcription support, document generation for PDF/DOCX/XLSX/PPTX formats, image transformation capabilities for existing images
✅ Binary File Protection (v0.1.4): Automatic blocking prevents text tools from accessing 40+ binary file types including images, videos, audio, archives, and Office documents, intelligent error messages guide users to appropriate specialized tools for binary content
✅ Crawl4AI Integration (v0.1.4): Intelligent web scraping with LLM-powered content extraction and customizable extraction patterns for structured data retrieval from websites
✅ Post-Evaluation Workflow (v0.1.3): Winning agents evaluate their own answers before submission with submit and restart capabilities, supports answer confirmation and orchestration restart with feedback across all backends
✅ Multimodal Understanding Tools (v0.1.3): Analyze images, transcribe audio, extract video frames, and process documents (PDF/DOCX/XLSX/PPTX) with structured JSON output, works across all backends via OpenAI GPT-4.1 integration
✅ Docker Sudo Mode (v0.1.3): Privileged command execution in Docker containers for system-level operations requiring elevated permissions
✅ Intelligent Planning Mode (v0.1.2): Automatic question analysis determining operation irreversibility via _analyze_question_irreversibility() in orchestrator, selective tool blocking with set_planning_mode_blocked_tools() and is_mcp_tool_blocked() methods, read-only MCP operations during coordination with write operations blocked, zero-configuration transparent operation, multi-workspace support
✅ Model Updates (v0.1.2): Claude 4.5 Haiku model claude-haiku-4-5-20251001, reorganized Claude model priorities with claude-sonnet-4-5-20250929 default, Grok web search fix with _add_grok_search_params() method for proper extra_body parameter handling
✅ Custom Tools System (v0.1.1): User-defined Python function registration using ToolManager class in massgen/tool/_manager.py, cross-backend support alongside MCP servers, builtin/MCP/custom tool categories with automatic discovery, 40+ examples in massgen/configs/tools/custom_tools/, voting sensitivity controls with three-tier quality system (lenient/balanced/strict), answer novelty detection preventing duplicates
✅ Backend Enhancements (v0.1.1): Gemini architecture refactoring with extracted MCP management (gemini_mcp_manager.py), tracking (gemini_trackers.py), and utilities, new capabilities registry in massgen/backend/capabilities.py documenting feature support across all backends
✅ PyPI Package Release (v0.1.0): Official distribution via pip install massgen with simplified installation, global massgen command accessible from any directory, comprehensive Sphinx documentation at docs.massgen.ai, interactive setup wizard with use case presets and API key management, enhanced CLI with @examples/ prefix for built-in configurations
✅ Docker Execution Mode (v0.0.32): Container-based isolation with secure command execution in isolated Docker containers preventing host filesystem access, persistent state management with packages and dependencies persisting across conversation turns, multi-agent support with dedicated isolated containers for each agent, configurable security with resource limits (CPU, memory), network isolation modes, and read-only volume mounts
✅ MCP Architecture Refactoring (v0.0.32): Simplified client with renamed MultiMCPClient to MCPClient reflecting streamlined architecture, code consolidation by removing deprecated modules and consolidating duplicate MCP protocol handling, improved maintainability with standardized type hints, enhanced error handling, and cleaner code organization
✅ Claude Code Docker Integration (v0.0.32): Automatic tool management with Bash tool automatically disabled in Docker mode routing commands through execute_command, MCP auto-permissions with automatic approval for MCP tools while preserving security validation, enhanced guidance with system messages preventing git repository confusion between host and container environments
✅ Universal Command Execution (v0.0.31): MCP-based execute_command tool works across Claude, Gemini, OpenAI, and Chat Completions providers, comprehensive security with permission management and command filtering, code execution in planning mode for safer coordination
✅ External Framework Integration (v0.0.31): Multi-agent conversations using external framework group chat patterns, smart speaker selection (automatic, round-robin, manual) powered by LLMs, enhanced adapter supporting native group chat coordination
✅ Audio & Video Generation (v0.0.31): Audio tools for text-to-speech and transcription, video generation using OpenAI's Sora-2 API, multimodal expansion beyond text and images
✅ Multimodal Support Extension (v0.0.30): Audio and video processing for Chat Completions and Claude backends (WAV, MP3, MP4, AVI, MOV, WEBM formats), flexible media input via local paths or URLs, extended base64 encoding for audio/video files, configurable file size limits
✅ Claude Agent SDK Migration (v0.0.30): Package migration from claude-code-sdk to claude-agent-sdk>=0.0.22, improved bash tool permission validation, enhanced system message handling
✅ Qwen API Integration (v0.0.30): Added Qwen API provider to Chat Completions ecosystem with QWEN_API_KEY support, video understanding configuration examples
✅ MCP Planning Mode (v0.0.29): Strategic planning coordination strategy for safer MCP tool usage, multi-backend support (Response API, Chat Completions, Gemini), agents plan without execution during coordination, 5 planning mode configurations
✅ File Operation Safety (v0.0.29): Read-before-delete enforcement with FileOperationTracker class, PathPermissionManager integration with operation tracking methods, enhanced file operation safety mechanisms
✅ External Framework Integration (v0.0.28): Adapter system for external agent frameworks with async execution, code execution in multiple environments (Local, Docker, Jupyter, YepCode), ready-to-use configurations for framework integration
✅ Multimodal Support - Image Processing (v0.0.27): New stream_chunk module for multimodal content, image generation and understanding capabilities, file upload and search for document Q&A, Claude Sonnet 4.5 support, enhanced workspace multimodal tools
✅ File Deletion and Workspace Management (v0.0.26): New MCP tools (delete_file, delete_files_batch, compare_directories, compare_files) for workspace cleanup and file comparison, consolidated _workspace_tools_server.py, enhanced path permission manager
✅ Protected Paths and File-Based Context Paths (v0.0.26): Protect specific files within write-permitted directories, grant access to individual files instead of entire directories
✅ Multi-Turn Filesystem Support (v0.0.25): Multi-turn conversation support with persistent context across turns, automatic .massgen directory structure, workspace snapshots and restoration, enhanced path permission system with smart exclusions, and comprehensive backend improvements
✅ SGLang Backend Integration (v0.0.25): Unified vLLM/SGLang backend with auto-detection, support for SGLang-specific parameters like separate_reasoning, and dual server support for mixed vLLM and SGLang deployments
✅ vLLM Backend Support (v0.0.24): Complete integration with vLLM for high-performance local model serving, POE provider support, GPT-5-Codex model recognition, backend utility modules refactoring, and comprehensive bug fixes including streaming chunk processing
✅ Backend Architecture Refactoring (v0.0.23): Major code consolidation with new base_with_mcp.py class reducing ~1,932 lines across backends, extracted formatter module for better code organization, and improved maintainability through unified MCP integration
✅ Workspace Copy Tools via MCP (v0.0.22): Seamless file copying capabilities between workspaces, configuration organization with hierarchical structure, and enhanced file operations for large-scale collaboration
✅ Grok MCP Integration (v0.0.21): Unified backend architecture with full MCP server support, filesystem capabilities through MCP servers, and enhanced configuration files
✅ Claude Backend MCP Support (v0.0.20): Extended MCP integration to Claude backend, full MCP protocol and filesystem support, robust error handling, and comprehensive documentation
✅ Comprehensive Coordination Tracking (v0.0.19): Complete coordination tracking and visualization system with event-based tracking, interactive coordination table display, and advanced debugging capabilities for multi-agent collaboration patterns
✅ Comprehensive MCP Integration (v0.0.18): Extended MCP to all Chat Completions backends (Cerebras AI, Together AI, Fireworks AI, Groq, Nebius AI Studio, OpenRouter), cross-provider function calling compatibility, 9 new MCP configuration examples
✅ OpenAI MCP Integration (v0.0.17): Extended MCP (Model Context Protocol) support to OpenAI backend with full tool discovery and execution capabilities for GPT models, unified MCP architecture across multiple backends, and enhanced debugging
✅ Unified Filesystem Support with MCP Integration (v0.0.16): Complete FilesystemManager class providing unified filesystem access for Gemini and Claude Code backends, with MCP-based operations for file manipulation and cross-agent collaboration
✅ MCP Integration Framework (v0.0.15): Complete MCP implementation for Gemini backend with multi-server support, circuit breaker patterns, and comprehensive security framework
✅ Enhanced Logging (v0.0.14): Improved logging system for better agents' answer debugging, new final answer directory structure, and detailed architecture documentation
✅ Unified Logging System (v0.0.13): Centralized logging infrastructure with debug mode and enhanced terminal display formatting
✅ Windows Platform Support (v0.0.13): Windows platform compatibility with improved path handling and process management
✅ Enhanced Claude Code Agent Context Sharing (v0.0.12): Claude Code agents now share workspace context by maintaining snapshots and temporary workspace in orchestrator's side
✅ Documentation Improvement (v0.0.12): Updated README with current features and improved setup instructions
✅ Custom System Messages (v0.0.11): Enhanced system message configuration and preservation with backend-specific system prompt customization
✅ Claude Code Backend Enhancements (v0.0.11): Improved integration with better system message handling, JSON response parsing, and coordination action descriptions
✅ Azure OpenAI Support (v0.0.10): Integration with Azure OpenAI services including GPT-4.1 and GPT-5-chat models with async streaming
✅ MCP (Model Context Protocol) Support (v0.0.9): Integration with MCP for advanced tool capabilities in Claude Code Agent, including Discord and Twitter integration
✅ Timeout Management System (v0.0.8): Orchestrator-level timeout with graceful fallback and enhanced error messages
✅ Local Model Support (v0.0.7): Complete LM Studio integration for running open-weight models locally with automatic server management
✅ GPT-5 Series Integration (v0.0.6): Support for OpenAI's GPT-5, GPT-5-mini, GPT-5-nano with advanced reasoning parameters
✅ Claude Code Integration (v0.0.5): Native Claude Code backend with streaming capabilities and tool support
✅ GLM-4.5 Model Support (v0.0.4): Integration with ZhipuAI's GLM-4.5 model family
✅ Foundation Architecture (v0.0.3): Complete multi-agent orchestration system with async streaming, builtin tools, and multi-backend support
✅ Extended Provider Ecosystem: Support for 15+ providers including Cerebras AI, Together AI, Fireworks AI, Groq, Nebius AI Studio, and OpenRouter
We welcome community contributions to achieve these goals.
Version 0.1.97 picks up the image/video edit work deferred from v0.1.86-v0.1.96 and continues multimodal provider-parity work: