AtlasMind is a VS Code extension that provides a multi-agent AI orchestrator with model routing, persistent project memory, and a skills registry. It turns VS Code's chat panel into an intelligent development assistant.

AtlasMind complements Copilot. While Copilot excels at inline completion, AtlasMind adds:

• Multiple agents with specialised roles
• Persistent project memory (SSOT) across sessions
• Multi-model routing across 8 providers with budget/speed control
• Autonomous project execution via /project
• 26 built-in skills including git, terminal, test runner, and code intelligence
• Custom skills with security scanning
• MCP server integration

Anthropic (Claude), OpenAI (GPT-4.1, o3), Google (Gemini), Mistral, DeepSeek, z.ai, GitHub Copilot, and any local model via OpenAI-compatible endpoints (Ollama, LM Studio, etc.).

The extension itself is free. LLM API usage is billed by each provider. Subscription providers (e.g. Copilot) and local models incur no additional cost. Use budget mode = cheap to minimize API spend.

---

• Ensure you're running VS Code ≥ 1.95.0
• Check that the extension is enabled in the Extensions panel
• Try reloading the window (Ctrl+Shift+P → Developer: Reload Window)

1. Ctrl+Shift+P → AtlasMind: Manage Model Providers
2. Click Set Key next to the provider
3. Keys are stored in VS Code SecretStorage (encrypted, never in settings files)

Yes — if you have GitHub Copilot active, AtlasMind can route through it. You can also connect a local model via Ollama or LM Studio (no API key needed).

1. Start Ollama or LM Studio with an OpenAI-compatible endpoint (e.g. http://localhost:11434/v1)
2. In the Model Provider panel, configure the Local provider with your endpoint URL
3. Models from the local server will appear in the Models tree view

---

The Single Source of Truth is a folder of Markdown files (project_memory/) that stores project knowledge: architecture, decisions, conventions, roadmap, and more. AtlasMind reads from it automatically to give the LLM project context.

• /bootstrap — creates the folder structure and scaffolds initial content
• /import — scans your workspace and auto-populates memory from README, configs, and manifests
• Freeform — ask AtlasMind to "remember" something

• Check that atlasmind.ssotPath points to the correct folder (default: project_memory)
• Ensure the files are valid Markdown
• Try /memory <keyword> to test retrieval

Absolutely. They're plain Markdown files. Edit them in VS Code like any other file. Changes are picked up on the next query.

---

• Budget mode (atlasmind.budgetMode): cheap → expensive determines which price tier of models are considered
• Speed mode (atlasmind.speedMode): fast → considered determines which speed tier
• Agent constraints: Set allowedModels on a custom agent to force specific models
• auto mode: Let the task profiler decide based on the request

• Budget mode may be set to balanced or expensive
• The task profiler may have detected a high-reasoning task (in auto mode)
• The agent may have allowedModels set to premium models only
• Check /cost for a breakdown

• Premium models (e.g. Claude Opus 4 via Copilot) cost 3× per request
• Set budget mode to cheap or balanced to prefer 1× models
• The router automatically conserves quota when below 30% remaining

---

When the estimated file impact exceeds the approval threshold (default: 12 files), AtlasMind requires explicit approval. This prevents unintended large-scale changes. Adjust with atlasmind.projectApprovalFileThreshold.

• Checkpoints allow rollback of that subtask's changes
• Other subtasks that don't depend on it continue
• The synthesis report includes the failure details
• You can re-run from the Project Run Center

Yes — be specific in your goal:

• "Refactor src/auth/ to use JWT" (scoped to a directory)
• "Add unit tests for the CostTracker class" (scoped to a class)

---

Ctrl+Shift+P → AtlasMind: Add Skill → Choose "Create from template" or "Import existing file". Custom skills are scanned for security issues before enablement.

Check the security scan results:

• Ctrl+Shift+P → AtlasMind: Configure Scanner Rules
• Right-click the skill in the sidebar → Show Scan Results
• Error-level rules (eval, child_process, hardcoded secrets) block enablement

Ctrl+Shift+P → AtlasMind: Manage MCP Servers → Add a server with stdio or HTTP transport. Tools auto-register as skills.

Only ~40 pre-approved commands are allowed. Write-capable commands (npm install, etc.) require atlasmind.allowTerminalWrite = true and explicit approval. See Tool Execution for the allow-list.

---

• Check provider health in the Model Provider panel
• Switch to speed = fast for quicker responses
• Use a local model for zero-latency inference
• Reduce chatSessionTurnLimit to carry less context

• Set budget = cheap
• Use Copilot (subscription) or local models
• Set costLimitUsd on custom agents
• Review spending with /cost