anthropic-sdk-typescript sdk-v1.9.0
- Source
- github.com
- Author
- anthropic-sdks-releaser[bot]Top Viber
- Date

Context
Most of version 1.9.0 of Anthropic's TypeScript SDK tightens its toolset, the ready-made bash, read, write, edit, glob and grep tools a developer runs on their own machine when Claude does work there (release notes). The SDK's guide lists these tools among the building blocks for serving Claude Managed Agents sessions from a self-hosted environment, the kind of setup behind Anthropic's self-hosted Managed Agents beta. The toolset's own documentation states the trust model: the file tools stay inside a working directory plus any allowed roots and are safe without a , while bash can do anything and should run inside one. Several of the release's fixes close ways around that boundary.
The clearest cases involve programs the toolset starts. The grep tool runs ripgrep (`rg`) when it is installed, and downloaded skills are unpacked with the system's tar or unzip. The SDK used to take those programs from any directory on the PATH, including the current directory and directories inside the working directory, where the agent's own write tool can create files. The tests for the grep change plant a fake `rg` there and check that the tool now passes over it, using the next `rg` on the PATH or its built-in search instead, and tar and unzip follow the same rule. A separate fix makes grep ignore ripgrep configuration files, whose options apply to every search; its test sets a configuration that adds `--follow`, which would let grep follow a symbolic link out of the working directory, and checks that a file on the other side is no longer found. These programs also no longer receive the process's `ANTHROPIC_*` environment variables, which the code says hold the API key, auth and per-work session token. The bash tool already withheld them.
Other fixes handle symbolic links more strictly. Listings of the /memories directory leave out links that point outside it, memory files are read only through the path that was checked, and a file tool path that passes through a link to a missing file is now fully resolved. Anthropic lists all of these as bug fixes. They change behavior only for apps that use the toolset, and the release notes name one cost: a host whose only tar or unzip sits inside the working directory or an allowed root can no longer download skills until one is installed elsewhere.
Two changes reach beyond the toolset. The SDK's lenient parser for tool input that streams in pieces used to skip text that is not JSON, such as a comment, and join numbers such as `12 3`, so it could return a value the model never wrote. A stream with such input now fails with an error, while valid and truncated JSON parse as before. Anthropic has also deprecated seven old beta header values, among them `prompt-caching-2024-07-31` and `token-counting-2024-11-01`. The API still accepts them, but a later SDK release removes them, so requests should stop sending them. Two other values, `files-api-2025-04-14` and `skills-2025-10-02`, select the beta shape of responses, and Anthropic says to follow the Files API and Skills migration guides before deleting those. The Python SDK's 2.9.0 carries the same API changes and similar toolset fixes.
- The lenient parser for streamed tool input used to skip non-JSON text such as comments and join numbers like 12 3, returning values the text did not contain. It now rejects such input and the stream fails; valid and truncated JSON parse as before.
- setupSkills now takes tar, unzip and helpers such as gzip only from PATH entries outside workdir and allowedRoots. A host whose only tar or unzip sits inside those directories cannot download skills until one is installed elsewhere.
- Extracted files now get mode 0644, or 0755 if any execute bit is set, inside 0700 directories, whatever the archive recorded. The system tar must accept --no-same-owner and --no-same-permissions.
- The grep tool's built-in search now stops reading a file at a NUL byte. Previously, a very large binary file with no line break could end the process with an uncaught RangeError.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- sandbox — An isolated environment where AI-generated code or agent actions run without being able to touch anything real.
- token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
- agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.
Checking sign-in…
Loading comments…





