Sweep AI

Generates code suggestions by indexing the entire project locally and predicting multiple tokens ahead using a custom-trained 'Tab model'. Operates within milliseconds by leveraging local codebase context rather than sending full context to remote APIs, enabling instantaneous suggestions as developers type. The indexing mechanism maintains awareness of code structure, definitions, and patterns across the entire project to inform predictions.

Generates code snippets, functions, or refactorings by retrieving relevant context from the indexed codebase and synthesizing new code that aligns with project patterns. Uses code search and definition resolution to understand existing implementations, then generates code that matches the project's style, dependencies, and architectural patterns. Operates through chat or inline prompts within the IDE.

Implements a flexible pricing model where autocomplete is unlimited on paid plans, but advanced features (code generation, chat, code review, web search) consume API credits. Free tier includes 1,000 autocompletes and $5 API credits; paid tiers ($10-60/month) include unlimited autocomplete and varying API credit allowances. Operates by tracking feature usage and deducting credits per request, with optional automatic top-up for continuous usage.

Analyzes code changes between branches by comparing diffs and providing structured review feedback on correctness, style, and potential issues. Operates by fetching the diff between two branches (typically feature branch vs. main) and applying code review logic to identify problems, suggest improvements, and flag risky patterns. Integrates with the IDE's diff viewer for inline feedback.

Enables the agent to search the web and fetch content from URLs to augment code generation and problem-solving. Introduced in v1.24, this capability allows Sweep to retrieve external documentation, API references, library examples, and Stack Overflow answers to inform code suggestions. Operates by parsing search queries, fetching relevant web content, and incorporating findings into the generation context.

Integrates with remote Model Context Protocol (MCP) servers to extend agent capabilities beyond built-in tools. Supports OAuth 2.0 and 2.1 authentication for secure server connections, allowing Sweep to invoke custom tools, access external services, and orchestrate multi-step workflows through standardized MCP protocol. Introduced in v1.27, this enables third-party tool integration without modifying core agent code.

Resolves code definitions and enables semantic search across the entire indexed project to understand code structure, dependencies, and relationships. Allows the agent to navigate from a symbol to its definition, find all usages, and understand the call graph — essential for context-aware code generation and refactoring. Operates by parsing code structure (likely using AST or language-specific parsers) and maintaining a searchable index of definitions.

Provides code completion suggestions with syntax highlighting and language-specific formatting, ensuring suggestions respect language grammar and conventions. Introduced in v1.26, this capability enhances autocomplete by rendering suggestions with proper syntax coloring and indentation, making suggestions more readable and reducing errors from malformed code. Operates by parsing the current language context and applying language-specific rendering rules.

Indexes the entire project locally for context-aware suggestions while maintaining strict privacy guarantees: code is never trained on, never retained by third parties, and can be disabled via Privacy Mode. Operates by building a local index of code structure and patterns without uploading raw code to remote servers. Compliant with SOC 2 standards and supports zero data retention policies for enterprise deployments.

Uses custom-trained LLMs optimized for code tasks ('Tab model' for autocomplete, custom models for generation) with claims of 'unmatched price, performance, and security' compared to generic LLMs. Operates by deploying proprietary models trained on code-specific datasets, avoiding reliance on general-purpose models like GPT-4. Model selection and configuration options not fully documented, but pricing suggests different models for different tasks.

Provides a chat interface integrated directly into JetBrains IDEs for conversational code discussion, problem-solving, and iterative refinement. Allows developers to ask questions about code, request explanations, and refine suggestions through natural language dialogue without leaving the IDE. Operates by maintaining conversation context and applying chat-based code generation to produce iterative improvements.