ms-agent

Call submit_research_task with your research query to get a task_id, poll with check_research_progress every few minutes, then call get_research_report when the status is completed. The async pattern avoids blocking the agent during the 20-60 minute run.

Use submit_code_genesis_task with a natural language query describing the project, choosing the standard 7-stage or simple 4-stage workflow. Poll with check_code_genesis_progress and retrieve the file tree with get_code_genesis_result. Docker is required for sandboxed execution.

The lsp_check_directory and lsp_update_and_check tools support TypeScript/JavaScript via typescript-language-server, Python via pyright, and Java via Eclipse JDT Language Server. The corresponding language server must be installed for each language.

The web_search tool supports arxiv, exa, and serpapi engines. Arxiv is the default and requires no API key, while exa needs EXA_API_KEY and serpapi needs SERPAPI_API_KEY. Set fetch_content=true to retrieve full page text.

replace_file_contents matches exact text instead of line numbers, making it concurrent-safe when multiple agents edit different parts of the same file. It also supports replacing a specific occurrence or all occurrences without rewriting unchanged content.

Video generation requires Python 3.10+, Node.js 16+ for Remotion animation rendering, FFmpeg on PATH, and API keys for an LLM provider and an image generation provider such as modelscope, dashscope, or google. A typical video takes about 20 minutes across 9 pipeline steps.