Download
One installer. No account. Your keys stay on your machine.
macOS and Linux builds are on the roadmap.
Install guide
Open hedes-studio-setup.exe. If SmartScreen shows, click “More info → Run anyway”.
Settings → AI → Providers. Paste a key or point to a local model.
Pick a model, choose an access mode, and go.
Changelog
Autonomous Bot Execution & Full Access Mode: Added intuitive Bot Access Mode Selector (Full access, Accept edits, Ask first, Plan only) directly into the bot chat composer and page header. Teammates can now run commands, browser actions, and file edits autonomously without stopping to ask permission on every step.
Global "All Bots: Full Access" Switch: One-click header toggle applies autonomous execution across every bot in the ecosystem simultaneously, eliminating repetitive approval dialogs for long-running workflows.
1-Click "Approve & Full access" Card Action: Pending approval cards now feature instant full-access elevation, allowing you to approve an action and grant autonomous rights to the bot in a single click.
Autonomous Process Watchdog & Non-Blocking Daemons: Continuous 5-second liveness monitoring across commands. Instant resolution for servers, preview builds, and public tunnels (Cloudflare Tunnel, localtunnel, ngrok) upon URL generation, completely eliminating execution hangs.
Port Conflict Auto-Resolution & Zombie Prevention: Dedicated tracking and clean termination of previous servers and tunnels per project, preventing port 4173/5173 collisions and 502 Bad Gateway proxy errors.
Interactive Prompt Auto-Bypass & Vite Tunnel Protocol: Automated detection of terminal prompts ((y/n), confirmations) and enforced Vite allowedHosts protocols, guaranteeing smooth remote link access without host blockades.
Embedded Default Local Agentic Model Engine: Guaranteed out-of-the-box lightweight (1.5B) agentic model running directly on standard laptops without requiring external Ollama, Python, or GPU setups. Silent background initialization and automated watchdog failover when cloud providers stall or rate limit.
Bots Ecosystem Supercharging: Persistent AI teammates now wield the full suite of Studio superpowers—project codebase inspection (read_project_file, search_project), specialized sub-agent delegation (run_subagent), active HEDES skills integration, and autonomous ReAct todo checklists.
Enhanced AI Voice & Assistant Chat: Real-time project code awareness and resilient spoken failover transitions, keeping hands-free and Claude-style assistant workflows uninterrupted.
Interactive Question & Decision Cards: Antigravity-grade interactive clarification UI with clickable option chips ((Recommended) badges) and custom write-in input, enabling fast single-click answers for architectural trade-offs.
AI Council for High-Performance Coding: Specialized technical council archetypes (Architecture Lead, Security Auditor, Performance Engineer, Implementation DX) and streamlined engineering roadmaps for code reviews, refactoring, and multi-agent consensus.
Local Agentic Watchdog & Automatic Failover: Resilient watchdog system that detects stuck primary cloud models (429 rate limits, 5xx server downtime, network stalls), informs the user with simple clear text, automatically switches to local lightweight models without lagging laptop hardware, and seamlessly restores primary cloud routing once recovered.
Immutable Native IDE Brain: Built-in default awareness of HEDES Studio's entire architecture—embedded web browser automation, multi-engine live preview, terminal execution, AST inspection, and automatic Git checkpoints.
Polyglot Engineering Models: Dedicated language-specific mental models for C/C++ (modern C17/C++20, CMake, RAII, memory bounds, compile-first), Python, Flutter/Dart, TypeScript/Web, Rust, and Go.
Autonomous Goal Engine ("Work Until Finished"): Expanded 12-round ReAct loop with Todo Checklist tracking (- [x] done, - [~] current, - [ ] next) that drives multi-step tasks continuously until complete and verified.
Specialized Sub-Agent Swarms: run_subagent enhanced with explicit roles (researcher, reviewer, debugger) for isolated deep analysis without context pollution.
One-Question Clarification Gate: If user requirements are fundamentally ambiguous, HEDES asks exactly ONE sharp clarifying question; otherwise it executes autonomously end-to-end.
Autonomous Pre-Flight Verification Gate: Written code immediately undergoes background syntax analysis and missing-dependency audits. Any detected issue is fed directly back into the ReAct loop so HEDES self-heals before concluding.
Accurate Thinking Duration: Thought cards display precise thinking and execution durations in real-time.
Live Public Tunnel Recognition: Instant detection and extraction of Cloudflare Tunnel (*.trycloudflare.com), localtunnel (*.loca.lt), ngrok, and pinggy URLs directly presented in clear agent responses without confusing links with UI buttons.
Input Generation Lock & Stop Control: Chat input textarea is locked during agent runs with “HEDES is working… Click Stop to interrupt” to prevent message collision, with prominent Stop button control.
Seamless ReAct Verification: Eliminated false-positive missing-edit nudge loops so command execution, build steps, and tunnel deployments resolve smoothly with clean completion summaries.
Antigravity ReAct Autonomous Loop: Full multi-step agent workflow where HEDES reasons explicitly inside expandable thoughts (“Thought for Xs >”), runs build and verification commands, explores file structures, observes real terminal exit codes and outputs, and concludes with clear execution summaries.
Clean Timeline Cards: Integrated in-chat status cards for thought inspection, command runs, and file exploration without polluting chat bubbles with raw internal prompts.
Tool token sanitization & deep-click details: ChatGLM and Qwen tool calls (<|tool_calls_section_begin|>) are parsed and executed cleanly. Large code dumps and setup files in bot messages are organized into collapsible deep-click disclosures.
Team coordinator delegation protocol: multi-bot delegation strictly routes through real message_bot tool calls. Coordination runways extended to 28 steps with graceful progress preservation on step limits.
Full Access mode for Bots: bots inherit global full access mode to run approved terminal commands and file edits without repeated prompts, while keeping safety floor guards active for destructive commands.
Bot queue resilience: answering questions from Atlas Planner or other bots immediately resolves the pending turn without freezing in queue. Added 1-click “Clear queue” and bot resume controls.
Team coordinator enhancements: delegating tasks in #team automatically unpauses paused teammates (e.g. Forge), with a new 1-click “Resume all” header action.
HEDES flagship models pinned to top: Devstral 2 (HEDES Pro), Kimi K2.5 (HEDES Max), and GLM 4.7 Flash (HEDES Fast) appear at the very top of the model selector with dedicated badge chips.
Cleaner chat interface: background workspace inspection actions now display as clean collapsible status cards rather than raw prompt text in speech bubbles. Model reasoning and thoughts are organized into a collapsible drawer.
Extended AI model timeout and filtered unsupported parameters to prevent 502 upstream errors and 400 validation failures during large code generation tasks.
Chat and bot message boxes grow with longer prompts. Focused fields keep a calm border. Chat and project conversations gain editable names; bot chats get task-based names.
Bots show ordered queue positions, active work, recent run logs, and HEDES model cost estimates. Stop also aborts active runs and their delegated work. Delegation waits for teammate results.
Studio retries once when a requested code edit produces only prose. Bot avatars now use consistent icons.
Chat keeps failed and interrupted turns, saves longer histories with images, and shows a quiet working animation. Expand activity when you want tool details.
Research bots execute supported tools, report model failures, and retain partial work. Council retries transient failures and preserves successful contributions.
Pro and Max show their actual plans. Owners can configure AI budgets, request rates, answer limits, and global spending caps. Updated dependencies and atomic AI budget reservations strengthen security.