Skip to content
/
Tier B A8 Cross-runtime harness

Mysti

mysti · DeepMyst/Mysti · ★ 1.1k · last commit 2026-03-11

Primitive shape 29 total
Commands 5 Skills 12 Subagents 12
00

Summary

Mysti — Summary

Mysti is a VS Code extension that orchestrates 12 AI coding agent CLIs (Claude Code, Codex, Gemini, GitHub Copilot, Cline, Cursor, OpenClaw, OpenCode, Qwen Code, Manus, Ollama, LocalAI) through a unified chat UI with "Brainstorm Mode" as its flagship feature: select any two of the 12 agents and have them collaborate using one of 5 strategies (Quick synthesis, Debate, Red-Team, Perspectives, Delphi). It also offers 16 developer personas, 12 toggleable behavioral skills (concise, TDD, auto-commit, etc.), autonomous mode with safety classifier (conservative/balanced/aggressive), @-mention routing to specific agents, context compaction, and conversation history. Distributed via the VS Code Marketplace (ext install DeepMyst.mysti); v0.4.0 with 360 automated vitest tests. With 1,056 stars and 4 contributors, Mysti is unique in this batch as a VS Code extension that orchestrates terminal-based agent CLIs rather than building its own agent loop. Compared to the seeds, Mysti is closest to BMAD-METHOD (multiple persona files, brainstorm mode) but runs inside VS Code as a panel rather than as Claude Code skills.

01

Overview

Mysti — Overview

Origin

Built by DeepMyst (github.com/DeepMyst). 4 contributors. 1,056 stars. Apache-2.0. v0.4.0. Last commit 2026-03-11.

Philosophy

"Wisdom of the crowd where the collective intelligence of several agents outperforms a single one."

"Single-model output shouldn't be trusted blindly. Multiple agents across different providers have different training, different strengths, and different blind spots. When any two work together, each AI catches edge cases the other might miss."

"It's like having a senior dev and a tech lead review your code—except they actually discuss it first."

Provider Philosophy

"Switch providers with one click. No lock-in."

Mysti's core bet: every major AI provider has unique strengths, and developers should be able to leverage multiple providers without switching tools. It supports 12 providers with authentic logos in the UI.

Brainstorm Mode Philosophy

"Want a second opinion? Enable Brainstorm Mode and let two AI agents tackle your problem together. Choose any 2 of 12 agents."

"When auto-convergence is enabled, the discussion exits early once agents reach consensus — saving time without sacrificing quality."

Requirements Philosophy

"Already paying for Claude, ChatGPT, Gemini, or GitHub Copilot? You're ready to go. Mysti works with your existing subscriptions—no additional costs!"

Mysti wraps existing CLI subscriptions rather than requiring its own API keys.

02

Architecture

Mysti — Architecture

Distribution

  • VS Code Marketplace: ext install DeepMyst.mysti or Ctrl+P + paste
  • URL: marketplace.visualstudio.com/items?itemName=DeepMyst.mysti

Technology Stack

  • TypeScript + webpack (compiled to dist/extension.js)
  • VS Code Extension API (webview, activity bar, commands, settings)
  • vitest for testing (360 automated tests)
  • No backend server — everything runs locally via spawned CLI processes

Directory Structure

src/
  extension.ts           — Extension entry point
  managers/              — Agent process management
  providers/
    ChatViewProvider.ts  — Webview panel provider
    MystiCodeLensProvider.ts
    MystiFileDecorationProvider.ts
    ProviderRegistry.ts  — 12 provider registrations
    base/                — Base provider class
    claude/              — Claude Code provider
    cline/               — Cline provider
    codex/               — Codex provider
    copilot/             — GitHub Copilot provider
    cursor/              — Cursor provider
    gemini/              — Gemini provider
    localai/             — LocalAI provider
    manus/               — Manus provider
    ollama/              — Ollama provider
    openclaw/            — OpenClaw provider
    opencode/            — OpenCode provider
    qwen/                — Qwen Code provider
  types.ts
  utils/
  webview/               — Chat UI (React webview)
docs/
  BRAINSTORM.md
  AUTONOMOUS-MODE.md
  MENTIONS.md
  COMPACTION.md
  PERSONAS-AND-SKILLS.md
tests/                   — vitest tests (360 total)
scripts/
  sync-agents.js
  fetch-stargazers.js

VS Code Activation

  • Activity bar sidebar: "Mysti" icon
  • Ctrl+Shift+M (Mac: Cmd+Shift+M) — open Mysti
  • Commands: mysti.openChat, mysti.newConversation, mysti.addToContext, mysti.clearContext, mysti.openInNewTab

Required Runtime

  • VS Code with one or more supported CLI tools installed:
    • Claude Code, Codex, Gemini CLI, GitHub Copilot CLI, Cline, Cursor, OpenClaw, OpenCode, Qwen Code, Ollama, LocalAI, Manus
03

Components

Mysti — Components

12 Supported Providers

Claude Code, Codex, Gemini, GitHub Copilot, Cline, Cursor, OpenClaw, OpenCode, Qwen Code, Manus, Ollama, LocalAI

5 Brainstorm Strategies

Strategy Roles Best For
Quick Direct synthesis Simple tasks, fast answers
Debate Critic vs Defender Architecture decisions
Red-Team Proposer vs Challenger Security reviews
Perspectives Risk Analyst vs Innovator Greenfield design
Delphi Facilitator vs Refiner Complex problems, consensus

16 Developer Personas

Architect, Debugger, Security-Minded, Performance Tuner, Prototyper, Refactorer, Full-Stack, DevOps, Mentor, Designer, and 6 more

12 Toggleable Behavioral Skills

Concise, Test-Driven, Auto-Commit, First Principles, Scope Discipline, and 7 more

VS Code Commands

Command Shortcut Description
mysti.openChat Ctrl+Shift+M Open chat panel
mysti.newConversation — New conversation
mysti.addToContext — Add file to context
mysti.clearContext — Clear context
mysti.openInNewTab — Open in new editor tab

Key Features

Feature Description
Brainstorm Mode Two-agent collaboration with convergence detection
@-mention routing @claude, @gemini in message routes to that provider
File mentions @filename adds transient file context
Autonomous mode Three safety levels with learning memory
Continuation modes Goal-based or task-queue for extended autonomous sessions
Audit trail Every autonomous decision logged for review
Context compaction Automatic at 75% usage; native /compact for Claude; client-side summarization for others
Smart auto-suggestions Personas + actions suggested based on message content
Quick actions Welcome screen one-click common tasks
Conversation history All sessions saved and accessible
Plan detection Auto-detects multiple implementation plans in agent response

Configuration Settings (mysti.*)

mysti.defaultProvider, mysti.brainstorm.agents, mysti.brainstorm.strategy, mysti.brainstorm.autoConverge, mysti.brainstorm.maxDiscussionRounds, mysti.autonomous.safetyMode, mysti.accessLevel, mysti.claudePath, mysti.codexPath, mysti.geminiPath, ...

05

Prompts

Mysti — Prompts

Mysti is a VS Code extension orchestrator; its prompts are embedded in the TypeScript source rather than in human-readable markdown files. The persona and strategy definitions are the primary prompt content.

Excerpt 1: Brainstorm Strategy Definitions (from README)

| Strategy | Roles | Best For |
|---|---|---|
| Quick | Direct synthesis | Simple tasks, fast answers |
| Debate | Critic vs Defender | Architecture decisions, trade-offs |
| Red-Team | Proposer vs Challenger | Security reviews, edge case discovery |
| Perspectives | Risk Analyst vs Innovator | Greenfield design, technology selection |
| Delphi | Facilitator vs Refiner | Complex problems, reaching consensus |

Prompting technique: Role-assignment strategy — each strategy assigns specific roles to each of two agents, shaping how they approach the same problem. "Critic vs Defender" creates adversarial dynamics; "Facilitator vs Refiner" creates collaborative consensus-building. This is a structured multi-agent prompting taxonomy.

Excerpt 2: 16 Personas (from README)

| Persona | Focus |
|---|---|
| Architect | System design, scalability, clean structure |
| Debugger | Root cause analysis, bug fixing |
| Security-Minded | Vulnerabilities, threat modeling |
| Performance Tuner | Optimization, profiling, latency |
| Prototyper | Quick iteration, PoCs |
| Refactorer | Code quality, maintainability |
| + 10 more... | Full-Stack, DevOps, Mentor, Designer... |

Prompting technique: Persona injection — selecting a persona changes the AI's system prompt or instruction framing before each message. Each persona focuses the agent on a specific dimension of code quality.

Excerpt 3: @-Mention System Description (from README)

@claude Review this code for security issues
@src/auth.ts @gemini Suggest performance improvements for this file
@claude Write tests, then @codex optimize them
  • File mentions: @filename adds transient context
  • Agent mentions: @agent routes tasks to that provider
  • Chaining: Later agents receive earlier agents' responses as context

Prompting technique: Context-routing DSL — inline @ syntax as a mini-language for routing tasks between agents and injecting file context. Chaining enables prompt output → input pipelines across agents.

09

Uniqueness

Mysti — Uniqueness

differs_from_seeds

Mysti is closest to BMAD-METHOD among the seeds: both support multiple named personas and multi-agent "brainstorm" workflows, and both are methodology-focused rather than toolserver-focused. The architectural delta: BMAD-METHOD is a skills pack for Claude Code that uses TOML/MD persona files and activates via Claude Code's skill system; Mysti is a VS Code extension that spawns CLI subprocesses for 12 different agent providers. BMAD's personas are "iron law" behavioral instructions injected into one agent's context; Mysti's personas inject role descriptions into a single-agent session while Brainstorm Mode uses two separate agents with different roles. Mysti's convergence detection (auto-exit when agents agree) has no equivalent in any seed.

Positioning

Mysti is the "universal VS Code orchestrator for the AI CLI ecosystem." It targets developers who have subscriptions to multiple AI tools and want to use them from one interface without switching tools. The 12-provider support means it can run Claude, Codex, Gemini, Copilot, Cline, Cursor, Qwen, Ollama, LocalAI, OpenClaw, OpenCode, and Manus — essentially all major coding agent CLIs as of 2026.

Observable Failure Modes

  1. Last commit 2026-03-11 — possible development slowdown (3 months since last push)
  2. VS Code only — no Cursor, JetBrains, or standalone support
  3. CLI process management complexity — spawning and managing 12 different CLI processes with different output formats, auth flows, and lifecycle behaviors is inherently fragile
  4. v0.4.0 stability fixes note — "18 stability fixes across brainstorm mode" in latest release; the multi-agent coordination has known edge cases
  5. 4 contributors — limited bandwidth for 12 provider adapters

What Is Genuinely Novel

  • 5-strategy Brainstorm taxonomy (Quick/Debate/Red-Team/Perspectives/Delphi) as a structured multi-agent collaboration framework
  • Convergence detection — heuristic-based auto-exit when agents reach consensus
  • @-mention chaining across providers in a single message (earlier agent's output becomes later agent's context)
  • Unified 12-provider chat interface with per-conversation provider switching
  • 16 persona + 12 skills combination space — 192 possible behavioral configurations
  • Windows fix for spawn EINVAL in v0.4.0 — demonstrates cross-platform CLI spawning complexity
04

Workflow

Mysti — Workflow

Basic Workflow

  1. Install Mysti from VS Code Marketplace
  2. Install at least one CLI tool (Claude Code, Codex, Gemini, etc.)
  3. Open Mysti: click icon in Activity Bar or Ctrl+Shift+M
  4. Select provider (single click)
  5. Type request → AI responds in chat panel
  6. Use @-mentions to route to specific agents inline

Brainstorm Mode Workflow

  1. Enable Brainstorm Mode in settings panel
  2. Choose two agents from 12 available
  3. Choose strategy (Quick/Debate/Red-Team/Perspectives/Delphi)
  4. Ask question → both agents process independently first
  5. Agents see each other's responses and build on them (configurable rounds, default 3)
  6. If auto-convergence enabled: exits early when agents agree
  7. Final synthesized response displayed

@-Mention Chaining Workflow

@claude Review this code for security issues
@src/auth.ts @gemini Suggest performance improvements for this file
@claude Write tests, then @codex optimize them

Later agents receive earlier agents' responses as context (chaining).

Autonomous Mode Workflow

  1. Enable autonomous mode (Conservative/Balanced/Aggressive)
  2. Safety Classifier runs on each proposed action:
    • Safe → auto-approve
    • Caution → mode-dependent (may ask or auto-approve)
    • Blocked → always deny
  3. Learning memory remembers permission preferences
  4. Continuation: Goal-based or task-queue for extended work
  5. Audit trail: every decision logged for review

Phase → Artifact Map

Phase Artifact
Start ~/.config/mysti (VS Code extension settings)
Conversation Conversation history (local storage)
Autonomous decision Audit trail log
Brainstorm Multi-round exchange + final synthesis
Compaction Summarized conversation history

Approval Gates

Gate Description
Permission prompts (autonomous mode) Caution-level actions ask user in ask-permission mode
Auto-commit Enabled/disabled via mysti.skills.autoCommit toggle
06

Memory Context

Mysti — Memory & Context

Memory Architecture

Mysti has a lightweight memory system focused on conversation history and learning permissions.

Conversation History

  • All conversations saved locally (VS Code extension storage)
  • Accessible from history panel
  • Searchable

Learning Memory (Autonomous Mode)

  • Remembers user permission preferences across sessions
  • When the same action is seen again, applies learned preference automatically
  • No external database — stored in VS Code extension state

Context Management

Context Window (Manual)

  • mysti.addToContext — add files to context
  • mysti.clearContext — clear all context
  • @filename mentions — add transient file context per message

Compaction (Automatic)

  • Triggers at 75% token usage threshold (default)
  • Claude Code: uses native /compact command
  • Other providers: client-side summarization (message summarization in TypeScript)
  • Per-panel tracking: each chat panel tracks usage independently

Cross-Session Handoff

Conversation history persists. Each conversation starts fresh but prior conversations are viewable.

State Files

VS Code extension stores state in VS Code's standard extension storage directory (not a user-visible path). No external files written to the project directory.

07

Orchestration

Mysti — Orchestration

Multi-Agent Support

Yes. Brainstorm Mode orchestrates exactly 2 agents simultaneously. @-mention chaining dispatches tasks sequentially to multiple agents.

Orchestration Pattern

Sequential for @-mention chaining (A → B, B receives A's output). Parallel-fan-out then synthesize for Brainstorm Mode (both agents see same prompt, produce independent responses, then see each other's).

Execution Mode

Interactive-loop (VS Code extension, user-driven).

Isolation Mechanism

Process-level — each CLI agent runs as a subprocess. No git worktrees or containers.

Multi-Model Support

Yes. Brainstorm Mode explicitly uses two different providers simultaneously. @-mentions route to specific providers. Provider selection per conversation is instant (one click).

Model role mapping in Brainstorm:

  • Agent 1: configured by mysti.brainstorm.agents[0]
  • Agent 2: configured by mysti.brainstorm.agents[1]

Consensus Mechanism

Convergence detection: Mysti tracks agent agreement and position stability. When autoConverge: true, exits discussion when agents reach consensus — without a formal protocol (heuristic-based).

Prompt Chaining

Yes. @-mention chaining explicitly chains outputs: @claude Write tests, then @codex optimize them — Codex receives Claude's test output as context.

Brainstorm Rounds

Configurable via mysti.brainstorm.maxDiscussionRounds (default 3). Each round: both agents see all prior contributions and add their own. Convergence guard prevents infinite loops.

Autonomous Mode Execution

Three safety levels control which agent actions are auto-approved:

  • Conservative: mostly blocks, asks for most things
  • Balanced: auto-approves safe operations
  • Aggressive: auto-approves safe + caution operations
08

Ui Cli Surface

Mysti — UI & CLI Surface

CLI Binary

None. Mysti is a VS Code extension, not a standalone CLI.

VS Code Extension UI

  • Type: VS Code Extension (webview + activity bar)
  • Activation: Activity bar icon + Ctrl+Shift+M
  • Tech stack: TypeScript + VS Code API + React webview + webpack

Chat Interface

  • Multi-provider chat with syntax highlighting and markdown
  • Mermaid diagram rendering
  • LaTeX support
  • Task list rendering with progress tracking
  • Persona toolbar (quick selection without opening settings)
  • @-mention autocomplete
  • Slash command menu
  • Auto-suggestions for personas/actions based on message

Settings Panel

  • Provider selection (12 providers with logos)
  • Brainstorm mode configuration (agents + strategy + max rounds)
  • Autonomous mode safety level
  • Persona selection (16 personas)
  • Skills toggles (12 skills)
  • Token budget settings
  • Access level (read-only / ask-permission / full-access)

Additional Views

  • Conversation history
  • Quick actions (welcome screen)
  • Code lens provider (contextual actions in editor)
  • File decoration provider

Observability

  • Autonomous mode audit trail (every decision logged for review)
  • Per-panel context usage tracking (triggers compaction)
  • Brainstorm round progress visible in UI

Cross-Tool Portability

Low-medium — VS Code extension only (no standalone CLI, no other IDE support). But within VS Code, supports 12 providers.

Related frameworks

same archetype · same primary tool · same memory type

claude-mem (thedotmack) ★ 78k
A8 Cross-runtime harness

Background worker service captures every tool call as an observation, AI-compresses sessions, and auto-injects relevant past…

pi (badlogic/earendil) ★ 55k
A8 Cross-runtime harness

A minimal, hackable, multi-provider terminal coding agent that adapts to your workflows via npm-installable TypeScript Extensions…

Agent Skills (Addy Osmani) ★ 46k
A8 Cross-runtime harness

Encodes senior-engineer software development lifecycle as 23 auto-routed skills and 7 slash commands for any AI coding agent.

wshobson/agents Plugin Marketplace ★ 36k
A8 Cross-runtime harness

Single Markdown source for 83 domain-specialized plugins that auto-generates idiomatic artifacts for five AI coding harnesses.

TabbyML/Tabby ★ 34k
A8 Cross-runtime harness

Self-hosted AI coding assistant server (alternative to GitHub Copilot) with admin dashboard, RAG-based completions, and multi-IDE…

Compound Engineering ★ 17k
A8 Cross-runtime harness

Make each unit of engineering work compound into easier future work via brainstorm→plan→execute→review→learn cycles.