Sanjeet Chatterjee develops Parselbox, an embeddable Python execution environment for AI agents. His work explores giving agents access to tools through code, and his professional biography identifies work on agents at Salesforce.[1][2][3]
Start with Parselbox – an embeddable Python sandbox for AI agents →
GitHub 
X Contributions
Parselbox exposes MCP servers, APIs and host capabilities as Python objects. An agent can combine them with loops, conditionals and local variables instead of returning each intermediate result through the conversation. A disk-backed workspace retains files and state between calls. Deno and Pyodide provide the execution layer, with explicit filesystem and network permissions rather than a container or virtual machine.[2]
Ideas
The project provides help, search and inspection operations so agents can discover capabilities as needed. Memory limits, timeouts, mounted directories and network allowlists configure its boundaries. Those controls still need to match the host capabilities supplied by the integrating application.[2]
Tools, writing & talks
Writing, research & talks
Start here
A Python execution layer that lets agents discover and compose MCP tools, APIs and host capabilities as code.
Begin with the architecture and permissions sections, then follow the Python API quick start. Notice which capabilities stay on the host and which filesystem and network permissions the runtime receives.
Developer and maintainer[1][2]Original source