Execute stateful Python or TypeScript code in a sandbox with capability-scoped side effects.
npx skills add https://github.com/ldclabs/langshell --skill langshell
Please help me install this Agent Skill. Skill: langshell Source: https://github.com/ldclabs/langshell/tree/main Command: npx skills add https://github.com/ldclabs/langshell --skill langshell
This Skill eliminates unsafe and over-permissioned tool use by letting AI agents run multi-step Python or TypeScript logic inside a persistent sandbox where all side effects are gated by explicitly registered capabilities.
Use case example: In an agent workflow that must transform data end-to-end, use LangShell to fetch approved JSON, filter and aggregate it, write a sanitized report into an authorized mount, and return a compact summary object.
Use the langshell skill to compute a safe, structured result and call only approved capabilities by writing Python or TypeScript code that assigns the final value to result.
High-intent search queries and answers about installing and using this skill.
To run sandboxed code for AI agents with controlled access, use a persistent execution environment that gates side effects through explicitly registered capabilities like read_text, write_text, and fetch_json. This enforces capability-scoped access for safe multi-step computation.
Capability-scoped execution is a sandboxing mechanism that requires AI agents to discover registered tools and comply with validation rules before performing side effects, ensuring deterministic multi-step computation with controlled access to files, HTTP endpoints, and databases.
Yes, you can maintain session state across multiple turns by using persistent sandboxed execution. This allows AI agents to keep variables, caches, and helper functions available throughout the same session for continuous multi-step data transformation.
You can validate agent tool calls and catch permission errors before execution by using a pre-run dry-run validation step. This process identifies unsupported features, unknown tools, and permission issues before any side effects occur in the sandbox.
The best way to return structured JSON results from sandboxed agent code is to assign the final computed value to a result variable, ensuring the output is JSON-serializable and produces a reliable RunResult object for downstream systems to consume.
Limitations of a persistent sandbox include strict enforcement of capability-scoped access, meaning agents cannot execute unregistered tools or perform unapproved side effects, and all results must adhere to serialization rules for structured result capture.
Knuckles-Team...
hatayamaFind Unity scene GameObjects by name, path, regex, component, tag, or layer.
I-OnlabsCoordinate multi-step agent workflows by streaming outputs between steps.
kiwamizamuraiBypass Python sandbox restrictions to execute system commands.
