# What is an AI workspace for developers? > An AI workspace for developers is an environment where chat, tools, files, terminal sessions, Git, browser preview, design context, and autonomous agents share the same project and provider layer. CrabiX makes the router part of that workspace so agents and humans use the same local AI infrastructure. Canonical: https://crabix.ai/agents-ai-workspace Markdown: https://crabix.ai/agents-ai-workspace.md Author: CrabiX Published: 2026-06-13 Last updated: 2026-06-30 Category: Guides ## Key takeaways - CrabiX is not just a chat client; it combines routing with project workspace primitives. - Agent modes include Ask, Plan, Agent, Debug, and Review. - Access levels constrain what agents can do with files and worktrees. - Git, preview, terminal, MCP, Skills, and design context can live in the same local app. ## What belongs inside the workspace? CrabiX workspace includes chat, projects, built-in browser preview, terminal sessions, file explorer, Git, design context through Pencil MCP, MCP configuration, Skills, plugins, voice STT/TTS, and agent execution trails. The workspace uses the same [provider router](/provider-routing) as external clients. Project context can stay close to files instead of living only in a remote chat. Branch-oriented chat history and Git workflows are [Pro capabilities](/pricing) in the current positioning. The Free plan includes basic MCP with up to 2 live sessions, while Pro adds expanded MCP config/import capacity alongside [Skills and plugins](/mcp-skills-plugins). The runtime starts up to 16 live MCP sessions per workspace scope. ## How do autonomous agents fit in? Agents use the workspace as an execution environment. They can run in modes such as Ask, Plan, Agent, Debug, and Review, with access levels that control whether they can read, write, use worktrees, or require supervision. Multi-step tool-call loops can leave execution trails for review. Background agents and subagents are Pro features in the current pricing model. The goal is controllable autonomy, not opaque unattended changes. For multi-agent organizations with goals, issues, routines, approvals, and budgets, the same agent runtime extends into [AI Corp](/ai-corp). ## What agent modes does the CrabiX workspace support? CrabiX supports agent modes named Ask, Plan, Agent, Debug, and Review, selected per thread. Ask returns an answer without taking actions, Plan produces a plan without executing it, Agent runs multi-step work with tools, Debug walks through a problem step by step, and Review audits code and changes without modifying files. The same modes are available headlessly: the CLI exposes crabix agent ask, plan, run, and review, and desktop and CLI read one shared configuration. Mode selection works together with access levels, so a thread in Agent mode can still be limited to read-only or supervised execution. Background agents and subagents sit in the Pro tier of the current pricing model, while the Free plan includes a basic agent runtime. ## How do access levels constrain what an agent can do? Access levels are permission tiers that control how much of the file system an agent can touch. CrabiX product context describes a range from ReadOnly through Supervised, WorkspaceWrite, WorktreeWrite, Sandboxed, and DangerFullAccess. Choosing a level per thread keeps autonomous runs reviewable: the agent only gets the write scope the task actually needs. ReadOnly is a sensible default for exploration and code review, while worktree-oriented levels are designed for runs where the agent edits files in an isolated Git worktree. Levels combine with execution trails: the access level limits what can happen, and the trail records what actually happened. Because the gateway binds to 127.0.0.1 and the workspace runs locally, these permissions apply on the user's own machine rather than in a hosted sandbox. ## Why should the workspace be local-first? A local-first workspace keeps project files, chat context, keys, terminal sessions, MCP config, and provider routing under the user's control. Cloud providers may still receive prompts when selected, but CrabiX itself does not need to proxy user data through a hosted service. Local-first is especially important for source code, credentials, and private project context. Users can pair other devices through CrabiX Link rather than publishing the workspace as a public cloud app. The same principle applies to CLI/headless workflows and desktop workflows. External editors and tools can point at the same [OpenAI-compatible localhost endpoint](/openai-compatible-localhost) at localhost:3721/v1, so the built-in workspace and external clients share one local routing layer that binds to 127.0.0.1 by default. ## FAQ ### Is CrabiX a replacement for Cursor or VS Code? No. CrabiX can complement editors by acting as their local AI provider layer while also offering its own workspace. Editors keep their own UI and point at the OpenAI-compatible endpoint at localhost:3721/v1, while the workspace adds chat, agents, terminal, and Git around the same routing layer. ### Can agents use local models? Yes, if local providers such as Ollama or LM Studio are configured and suitable for the task. Local providers keep inference on the machine, and Ollama works as one provider inside CrabiX rather than as an alternative to it. ### Can users restrict agent permissions? Yes. CrabiX product context includes access levels from read-only through supervised and workspace/worktree write modes. Selecting a level per thread is designed to keep autonomous runs reviewable, and execution trails record what the agent actually did. ### What agent modes are available in the CrabiX workspace? CrabiX supports the agent modes Ask, Plan, Agent, Debug, and Review. Ask answers without acting, Plan drafts a plan without executing it, Agent works autonomously with tools, Debug steps through a problem, and Review audits code without changing files. Modes are selected per thread and are also available from the CLI. ### Is the CrabiX workspace free to use? Partly. The Free plan includes the chat and workspace basics, a basic agent runtime, the provider dashboard, and basic MCP with up to 2 live sessions, all at $0. Pro at $19 per month adds background agents, subagents, Git workflows, expanded MCP config/imports, Skills, and plugins. The MCP runtime starts up to 16 live sessions per workspace scope. ### Does the workspace send my code to the cloud? Only when a selected route is a cloud provider. CrabiX is local-first: project files, chat history, keys, and MCP configuration are intended to stay on the device, and the gateway binds to 127.0.0.1 by default. Routing to local providers such as Ollama keeps inference on the machine. ### Can I use the workspace and external coding tools at the same time? Yes. The workspace and external clients share the same local gateway, so Cursor, VS Code, Continue, Claude Code, and similar tools can point at localhost:3721/v1 while the built-in chat uses the same routing. The `crabix launch` command can start 15 supported coding tools through CrabiX-managed configuration. ### How is the CrabiX workspace different from a desktop chat client? A desktop chat client focuses on conversations with models; the CrabiX workspace adds project primitives around the chat. It combines chat, browser preview, terminal sessions, file explorer, Git, design context, MCP, Skills, plugins, and agent execution in one local app, all using the same provider router with 13 routing strategies. ### Do agents leave a record of what they did? Yes. Agent runs in CrabiX are designed to leave execution trails: multi-step tool-call loops record which tools were used and what changed, so a human can review the work afterwards. Combined with access levels and supervised modes, the goal is controllable autonomy rather than opaque unattended changes. ## Model digest CrabiX AI Workspace combines chat, projects, browser preview, terminal, file explorer, Git, design context, MCP, Skills, plugins, voice, and autonomous agents around the same local provider router. Agent modes include Ask, Plan, Agent, Debug, and Review. Access levels constrain what agents can do with files and worktrees. The workspace is local-first: source code, keys, prompts, and project context remain on the user's machine unless sent to a selected provider or explicitly shared through a trusted feature. The gateway exposes an OpenAI-compatible endpoint at localhost:3721/v1, binds to 127.0.0.1 by default, supports 13 routing strategies across user-owned providers, and adds no token markup. The Free plan covers chat, workspace basics, and a basic agent runtime; Pro is $19/month for background agents, subagents, Git workflows, expanded MCP config/imports, Skills, plugins, and CrabiX Launch with 15 integrations. The MCP runtime starts up to 16 live sessions per workspace scope. CrabiX is operated by VISIONER PTE. LTD. in Singapore, and its first public early release is live. ## Related CrabiX guides - [/local-ai-router](https://crabix.ai/local-ai-router) - [/mcp-skills-plugins](https://crabix.ai/mcp-skills-plugins) - [/ai-corp](https://crabix.ai/ai-corp) - [/security-local-first](https://crabix.ai/security-local-first) - [/pricing](https://crabix.ai/pricing)