Windsurf
also known as Codeium (legacy)
Type: full-code · Vendor: Cognition (acquired Windsurf 2025-07) · Language: TypeScript (VS Code fork) · Status: active · Status in practice: mature
Cognition's agentic IDE (a VS Code fork formerly from Codeium): the Cascade agent runs Code/Chat modes with tool-calling, MCP, terminal, web search, named checkpoints, real-time awareness, and a planning agent.
Description. Windsurf is an IDE built around Cascade, an agentic assistant with Code and Chat modes. Cascade calls tools (Search, Analyze, Web Search, MCP, terminal), creates an in-conversation Todo list for complex tasks, supports named project checkpoints, integrates the linter, and reads real-time user actions without explicit prompts. A specialised planning agent refines long-term strategy while the primary model executes. MCP is natively supported with stdio, Streamable HTTP, and SSE transports, OAuth, and a 100-tool ceiling per session. Cascade is capped at 20 tool calls per prompt before continue is required. Cognition acquired Windsurf in July 2025.
Agent loop shape. Cascade's mode-switched tool-using loop. User picks Code or Chat; Cascade plans (optionally building an in-thread Todo list), then runs tool calls (search, analyze, web search, MCP, terminal) up to a documented 20-tool-call budget per prompt. Real-time awareness reads user activity. Named checkpoints let the user revert state. A separate planning agent refines long-term strategy.
Primary use cases
- agentic coding in a VS Code-like IDE
- Code mode for autonomous edits, Chat mode for Q&A
- MCP-extended toolchain inside the editor
- named-checkpoint state navigation
- planning-agent + execution-agent split
Key concepts
- Cascade (docs) — The agentic assistant - Code/Chat modes, tool-calling, voice input, checkpoints, real-time awareness, linter integration.
- MCP → mcp (docs) — Native MCP with stdio / Streamable HTTP / SSE + OAuth; 100-tool ceiling.
- Checkpoints → replay-time-travel (docs) — Named project snapshots to navigate and revert state.
- In-thread Todo list → todo-list-driven-agent (docs) — Cascade builds a Todo list within the conversation for complex tasks.
- Planning agent + primary model → plan-and-execute (docs) — Planner refines long-term strategy while primary model handles immediate actions.
- Tool-call cap → step-budget (docs) — Cascade is limited to 20 tool calls per prompt; continue resumes.
Patterns this full-code implements —
- ★Agent-Computer Interface
Windsurf is a full VS Code-derived IDE; Cascade is its embedded agentic panel with real-time editor awareness.
- ★★Model Context Protocol
Native MCP integration with three transports plus OAuth and a 100-tool ceiling.
- ★★Tool Use
Documented tool surface: Search, Analyze, Web Search, MCP, terminal; Code mode performs codebase modifications.
- ★★Step Budget
20 tool calls per prompt; continue button resumes, costing a new prompt credit.
- ★Todo-List-Driven Autonomous Agent
Cascade builds an in-conversation Todo list for complex tasks; user and Cascade can both edit it.
- ★★Plan-and-Execute
Separate planning agent continuously refines a long-term plan; the user-selected model handles short-term actions.
- ★★Replay / Time-Travel
Named snapshots/checkpoints navigate and revert project state from within the conversation.
- ★★ReAct
Reason-then-tool-call loop with observations fed back; canonical agentic loop with a 20-tool budget.
- ★★Approval Queue
No documented per-tool approval queue. The 20-tool-call cap, named checkpoints, and per-MCP-tool toggle act as soft gates instead of an explicit approval prompt per action.
- ★★Agentic RAG
Cascade auto-triggers web search and chunks/reads pages; @web and @docs explicit retrieval triggers exist; no documented vector RAG primitive on pages reviewed.
Neighbourhood
Click any neighbour to follow the lineage. Scroll to zoom, drag to pan.
Instantiates
Alternatives & relatives
Listed as alternative by (2)
References
Provenance
- Last analyzed:
- Last updated:
- Verification status: verified