AtlasMind (Beta)

_{Marketplace badge shows the latest published Marketplace version. The source version for this branch lives in package.json.}

_{AtlasMind remains in Beta until version 1.0.0, even though Marketplace publishes now ship on the standard release channel.}

AI coding inside VS Code, with model choice, project memory, approvals, and costs you can actually control.

AtlasMind is a VS Code extension for developers who want AI help without giving up control. It routes work across your models, stores project knowledge in plain Markdown, and keeps tool use, approvals, and spend visible instead of buried.

For newer developers, AtlasMind gives you guided entry points like /bootstrap, /import, and /project. /bootstrap now runs a skippable intake through Atlas chat so the same answers can seed SSOT memory, ideation defaults, project-scoped personality defaults, routing settings, and GitHub-ready planning artifacts. It also reuses future-looking details when they are provided early instead of asking the same question twice, including whether a project already has an online repo and where a new one should live if it does not. For experienced developers, it gives you multi-model routing, persistent SSOT memory, MCP extensibility, local-model support, and audit-friendly execution.

AtlasMind defaults to safety and evidence over blind autonomy. Its project workflow is built around safety-first execution, approval-aware changes, and red/green TDD-style autonomous delivery where implementation is expected to follow a visible failing-signal path instead of skipping straight to unchecked code edits.

Why AtlasMind

• Stay in standard VS Code: no custom fork and no browser-only workflow.
• Use the models you want: route across Anthropic, Claude CLI (Beta), OpenAI, Gemini, Azure OpenAI, Bedrock, Copilot, local OpenAI-compatible endpoints, and more.
• Keep project context: AtlasMind stores durable project memory in project_memory/ so architecture and decisions survive past one chat session, and the guided bootstrap can pre-seed that memory before the first implementation task.
• Start from safety: approval gates, verification hooks, memory scanning, and explicit execution controls are built in from the start.
• Favor red/green development: AtlasMind is designed to support tests-first autonomous delivery instead of opaque "trust me" code generation.
• Get real execution controls: approval gates, cost tracking, run history, checkpoints, and verification hooks are built in.
• Extend it cleanly: AtlasMind ships with 32 built-in skills and can grow through custom skills, MCP servers, and the shared runtime plugin surface.

30-Second Start

1. Install AtlasMind from the VS Code Marketplace.
2. Run AtlasMind: Manage Model Providers and configure one provider.
3. In chat, run @atlas /bootstrap for a new repo or @atlas /import for an existing one. The bootstrap flow can silently capture project brief, audience, budget, timeline, stack, tooling, repo-hosting intent, and personality-relevant project defaults, and it will carry forward details that were provided out of order.
4. Ask a real task, or run @atlas /project <goal> for a larger change.

That is enough to get productive. AtlasMind stores provider credentials in VS Code SecretStorage and loads project memory from the configured SSOT path or the default project_memory/ folder.

What You Can Do

Why Teams Pick It

• Project memory lives in plain files that can be reviewed, committed, and shared.
• Model choice is not locked to one vendor or one pricing model.
• Approvals, verification, and webhook hooks make autonomous work easier to trust.
• The CLI reuses the same orchestration model for headless or CI-style workflows.

Quick Comparison

Core Surfaces

• Chat and slash commands: @atlas, /bootstrap, /import, /project, /runs, /agents, /skills, /memory, /cost, /voice, /vision
• Guided bootstrap intake: /bootstrap can ask skippable product, team, timeline, budget, audience, stack, integration, and repo-hosting questions, infer future answers from earlier freeform responses, then seed project_soul.md, a project brief, ideation defaults, a repository plan, project-scoped Personality Profile defaults, roadmap prompts, workspace routing defaults, and GitHub-ready planning files.
• Command Palette: top-level AtlasMind surfaces such as Settings, Personality Profile, Model Providers, Agents, MCP Servers, Import VS Code MCP Servers, Project Dashboard, Project Ideation, Project Run Center, Voice, Vision, Cost, and Collapse All Sidebar Trees
• MCP import bridge: AtlasMind can scan the current VS Code profile mcp.json plus workspace .vscode/mcp.json files and copy compatible stdio or http servers into AtlasMind's own MCP registry, which is useful when a tool already works in Copilot chat but has not been registered in Atlas yet.
• Sidebar Home: the AtlasMind sidebar starts with a composite Home surface that groups quick actions, recent sessions, recent autonomous runs, and workspace status into internal accordion sections with remembered manual heights
• Personality Profile inputs: every prompt combines an editable freeform answer with quick-fill presets, can be saved either as a global default or as a project-specific override, and keeps editor-only load actions separate from the destructive clear-project-override action
• Live settings shortcuts: the Personality Profile live-settings tiles open the matching Atlas settings page so you can jump straight into models, chat, safety, or project configuration from the profile panel
• Sidebar home and actions: the top Home surface opens major AtlasMind workspaces and summarizes recent activity, while the lower views keep local actions for Agents, Skills, Sessions, Memory, Models, and MCP Servers. The Models tree keeps friendly names as the primary label but now shows exact model slugs inline whenever one provider exposes multiple variants with the same display name.
• Managed terminal chat launches: the shared Atlas chat surface can run bounded shell-integrated terminal commands through aliases such as @tps, @tpowershell, @tpwsh, @tgit, @tbash, and @tcmd, stream the output into the thread, and optionally let AtlasMind request one approval-gated follow-up command in the same session before summarizing the result. Profile- or remote-backed terminals like JavaScript Debug Terminal and Azure Cloud Shell are not wired into this managed runner yet, so those aliases currently return explicit guidance instead of silently failing.
• Live steering during responses: while AtlasMind is still responding in the shared chat surface, the composer remains editable so you can switch the send mode to Steer, submit a redirecting prompt, and have AtlasMind stop the active turn and continue immediately with the new steering instruction.

Voice

The Voice Panel uses the webview runtime's Web Speech APIs for browser-side STT and fallback TTS, with optional ElevenLabs server-side TTS when configured. AtlasMind now persists preferred microphone and speaker ids in workspace settings, applies output routing to ElevenLabs audio when the runtime supports setSinkId, and explicitly marks where browser speech still follows the default OS or browser device. OS-native speech backends are not wired in yet.

Configuration

AtlasMind's routing controls live under the atlasmind.* settings namespace. Alongside budgetMode, speedMode, and feedbackRoutingWeight, specialist routing can now be tuned per domain with atlasmind.specialistRoutingOverrides when a workspace needs to pin research, visual-analysis, voice, robotics, simulation, or media-generation requests to a preferred provider or fallback workflow surface. Local routing can also aggregate multiple labeled OpenAI-compatible endpoints through atlasmind.localOpenAiEndpoints, so one workspace can keep engines such as Ollama and LM Studio available together while AtlasMind still shows which endpoint owns each local model. The full settings reference lives in docs/configuration.md.

Detailed command and action reference lives in wiki/Chat-Commands.md.

Key Files

• src/providers/claude-cli.ts: Beta provider adapter that reuses a locally installed Claude CLI login through constrained print-mode execution.
• src/skills/dockerCli.ts: Strict Docker and Docker Compose skill for container inspection and controlled lifecycle operations.
• src/runtime/core.ts and src/extension.ts: shared provider seeding and extension-host registration for routed model backends.
• src/views/modelProviderPanel.ts: provider setup UX, readiness checks, and Beta labeling for model providers.

Documentation

• Getting Started
• Architecture
• Model Routing
• Memory System
• Agents & Skills
• Configuration
• Development Guide
• CLI
• Comparison
• Wiki Home

Contributing

See CONTRIBUTING.md for development setup, conventions, and extension points.

Local VSIX packaging uses the checked-in .vscodeignore to strip workspace-only artifacts such as project_memory/, wiki/, generated .vsix files, local Vitest JSON reports, and assistant instruction folders from test packages.

Support

AtlasMind is open source under MIT. If it saves you time, support development through GitHub Sponsors.

License

MIT — see LICENSE