MS Agent
An Apache 2.0 Python agent framework with a self-hosted browser workspace and local code execution. Default model requests use ModelScope's API.

MS Agent is an open source Python agent framework with a self-hosted web server for working on local projects. It's for developers building research assistants and coding workflows that continue across sessions. Code execution and project file work run locally; the default model integration uses ModelScope's API and requires an API key and internet access.
The browser workspace combines model chat with tool activity, file browsing and editing, and management of skills and MCP tools. Terminal interfaces and a Python SDK share the same agent execution logic. Other model integrations include OpenAI Generic and Anthropic; model requests go to the configured service independently of where the workspace runs.
Agents preserve session records, compress older context and reuse project memory in later tasks. Scheduling supports recurring and one-time jobs. A lead agent can delegate work to specialists, or developers can define stages and dependencies with separate models and tools for each role. Permission checks and hooks support plan validation, business rules and human intervention.
Built-in applications include Agentic Insight, which saves research artifacts to the filesystem and ties report claims to evidence, and CodeGenesis, which generates software and refines it using diagnostics and runtime feedback. DocResearch reads PDF, TXT, PPT and DOCX material and exports illustrated reports.
The framework uses the Apache 2.0 license. Its Skill Evolution workflow tests proposed skill revisions against a validation set and accepts them only when scores improve.