Install Mysti · Get started · 2.0 BETA release notes · Contribute
An open-source coding workspace that brings your agents together inside VS Code.
Ask Claude Code and Codex for independent opinions. Give a specialist a specific task. Hand completed work to another agent for review. Keep their responses, context, permissions, and progress in one conversation.
Mysti connects the coding tools you choose—CLI agents, local model servers, and the DeepMyst-powered Mysti coordinator. Use one agent when that's enough; bring in more when another perspective helps.
2.0 is a beta. This checkout prepares the 2.0.0 pre-release package. The Marketplace badge shows the currently published version. See release status and limitations before adopting it for important work.
16 provider adapters + the Mysti coordinator · Parallel opinions · Cross-agent handoffs · Connected tools · Proactive context
The recordings below use the shipped webviews with deterministic sample responses. They demonstrate interaction and presentation; they are not live provider runs or performance benchmarks. Capture details and video downloads.
An answer is a starting point. On answers with a usage or session receipt, click Second opinion, choose another configured agent, and Mysti sends it the original question. Compare separately attributed answers in the same conversation; use Brainstorm when you want a discussion and synthesis.
@claude @codex What are your opinions on adding a cache to this API?
Explicit tags determine who does the work. Independent read-only assignments run concurrently within the configured limit. Each agent has its own live card and attributed answer—even if one of them is already your selected provider.
A failed participant stays visible. Mysti does not invent its answer or quietly ask the other agent to impersonate it.
@claude Design a retry policy, then @codex review it for edge cases
Use then to make the next assignment depend on the previous result. Claude designs the policy; Codex receives that result to review. Progress cards show the current step and who runs next. A failed dependency blocks its dependent step.
Add roles such as @claude:critic or @codex:reviewer to make the perspective explicit. Independent advisory work can run in parallel; work that may edit a shared workspace runs serially. Assignment and workflow guide.
Choose from 16 registered provider adapters, plus the Mysti coordinator. Move between coding CLIs, local model servers, and API providers from the composer. Your draft stays in place, and model controls follow the selected provider's capabilities.
See all providers and setup requirements.
Choose the model, adjust effort, and toggle Ultracode where the provider supports it. The model menu and action menu share the same settings. Dictation adds text to your draft for review before you send.
Connect apps and MCP servers through DeepMyst's Composio and Smithery catalogs. Bring code, team discussions, email, and calendar context within reach of your agents through the tools you authorize. Browse and authorize in DeepMyst, then manage connections from Mysti.
The recording shows sample authorized connections, not a live OAuth flow. Catalog availability and tool permissions depend on the service and your account; local CLI access is opt-in for supported adapters. A connected app is not automatically a Proactive monitoring source.
Proactive brings selected local Git changes and DeepMyst-connected GitHub/Slack evidence into an inbox. The Before you start view helps you inspect relevant evidence for a responsibility, including source links and freshness information.
Cloud monitoring requires DeepMyst sign-in, a supported connection, and an explicitly configured responsibility. Coverage is bounded; evidence does not prove who owns a task. Proactive setup and limits.
Choose two configured agents and a discussion strategy. Quick collects independent analyses and synthesizes a response; Debate, Red Team, Perspectives and Delphi provide other ways to examine a proposal. Keep the individual reasoning visible alongside the synthesis.
| Feature | Why it matters | Explore |
|---|---|---|
| Brainstorm | Let two agents analyze, debate or challenge a proposal, then synthesize a response. | Discussion strategies |
| Canvas + Visual Test | Iterate on visual artifacts and inspect a running app alongside the code. | Visual workflows |
| Personas, skills and roles | Reuse project guidance and specialist perspectives instead of repeating instructions. | Team practices |
| Checkpoints + visible approvals | Inspect requested actions and recover supported local changes as you iterate. Checkpoints do not undo external actions. | Access and lifecycle |
| Dictation + context tools | Speak a draft, mention files, or attach diagnostics and Git context before sending. | Composer and context |
| When you need… | Use… | What happens |
|---|---|---|
| A direct answer or implementation | A selected provider, or @claude / @codex |
The named agent handles the request with your access settings. |
| Independent perspectives | @claude @codex with a question or review request |
Read-only assignments run concurrently; results remain separately attributed. |
| An ordered handoff | @claude Write the parser, then @codex review it |
The reviewer receives the completed result as reference material. A failed dependency blocks the handoff. |
| A defined perspective | @claude:critic @codex:reviewer |
Catalog roles shape each response and restrict its access. |
| A structured discussion | Brainstorm | Choose a team and discussion strategy, then follow its rounds and synthesis. |
| A coordinating agent | Mysti | A DeepMyst-backed coordinator can answer, use configured tools, and delegate through permission gates. |
| Visual iteration | Canvas and Visual Test | Work with visual artifacts and inspect a running app; browser setup may be required. |
| Project-specific practices | Personas, skills, and roles | Reuse your team's guidance with explicit trust and access boundaries. |
Parallelism follows the work: independent advisory requests may overlap; writers share a workspace and run serially. Use then for an explicit dependency. Assignment syntax and troubleshooting.
- Install Mysti from the VS Code Marketplace, or install the reviewed beta VSIX with Extensions → Install from VSIX. Choose the pre-release channel when 2.0 BETA is published.
- Open a trusted local workspace. Mysti supports VS Code 1.86 or newer on Windows, macOS, and Linux. Native voice support and provider CLIs have their own host requirements.
- Open Mysti and choose your agent. The setup screen checks installation and authentication and provides the supported install, upgrade, or sign-in action. Mysti itself uses your DeepMyst account; CLI agents use their own credentials.
- Start with a small task. Try
@claude Explain the main entry point, then add@codexfor a second opinion when both are configured.
You only need one configured agent to begin. Provider subscriptions, API access, usage charges, and supported models vary. Getting started · Provider setup.
| Provider family | Available adapters |
|---|---|
| Coding agents | Claude Code, OpenAI Codex, Google Gemini CLI, GitHub Copilot CLI, Cline, Cursor |
| Additional coding tools | OpenClaw, OpenCode, Qwen Code, Hermes, Continue, Kimi Code |
| Local / API endpoints | Ollama, LocalAI, OpenRouter, MiniMax |
| Mysti coordinator | DeepMyst account, with optional configured local execution and delegation |
The UI follows each adapter's capabilities. Model availability, effort, Ultracode, attachments, native approvals, and authentication are provider-specific. An adapter being included is not a claim that every provider/version/account combination has been live-tested. Compatibility and setup.
- Context where it belongs: mention workspace files, inspect diagnostics with
@problems, or add a Git summary with@git. - Visible authority: read-only and approval modes, agent-specific tool activity, and cancellation across active assignments.
- Continuity: conversation history, context compaction, and workspace checkpoints for supported local changes.
- Explicit connections: choose the DeepMyst tools and sources used by connected features. Cloud responsibilities continue on the backend; local Git watches need the editor running.
Provider requests transmit the prompt and supplied context to the configured provider. DeepMyst features use your connected account. Telemetry follows VS Code's telemetry setting; see the security policy and feature-specific guides. A local checkpoint cannot undo an external action.
| Guide | What it covers |
|---|---|
| Start here | Install, configure, send your first task, troubleshoot |
| Agent assignments | Tags, roles, parallel reviews, dependent handoffs |
| Providers | Installation, authentication, models, capability differences |
| Feature guide | Chat, menus, dictation, Canvas, permissions and more |
| Brainstorm | Team discussion and synthesis |
| Personas and skills | Reusable project practices |
| Proactive | Responsibilities, evidence, notification and coverage limits |
| Architecture | Host, webview, routing and provider boundaries |
| Release notes | What's new, migration, known limitations |
| Contributing | Development, meaningful tests, demos and release work |
Translations: العربية · 中文 · 日本語 · 한국어 · Español · Português · Deutsch · Français · Türkçe · Русский. Community translations may describe an earlier release; this README is the current 2.0 BETA reference.
Mysti is Apache-2.0 licensed. Useful contributions include reproducible routing bugs, native-provider compatibility checks, accessible UI improvements, and tested documentation corrections.
npm ci
npm run typecheck
npm test
npm run watchUse the Node version in .nvmrc; press F5 to launch the extension development host. See CONTRIBUTING.md for provider contracts, browser tests, and contribution expectations.
Report a bug · Discuss an idea · View the roadmap
Thanks to everyone who has helped make Mysti better!
Want to join them? Check out the contribution guide section below.
Apache License 2.0. Built by DeepMyst and the Mysti community. Provider names and trademarks belong to their respective owners.