# Status and compatibility ## Version and requirements | Item | Value | | --- | --- | | Version | 0.1.1 | | License | MIT | | Python | Python 3.11 or newer | | Dependencies | `google-cloud-storage>=2.18` | | `mcp` extra | `mcp>=2.2,<3`, for `wenchang.mcp`. Targets `mcp.server.mcpserver.MCPServer` only; FastMCP and mcp 1.x servers do not work. | | `testing` extra | `pytest>=8.3`, for the conformance suites in `wenchang.testing`. | wenchang is installed from git by release tag, for example `pip install "wenchang[mcp] @ git+https://github.com/msiebert/wenchang@v0.1.1"`. It is not published on PyPI. The base package works without the `mcp` extra; see [Integrate without MCP](integrate-without-mcp.md). ## Releases and versioning Releases are git tags named `vX.Y.Z`, each with a GitHub Release. There is no PyPI release. Changes are recorded in [`CHANGELOG.md`](https://github.com/msiebert/wenchang/blob/main/CHANGELOG.md) in the Keep a Changelog format, and the release process is in [`docs/releasing.md`](https://github.com/msiebert/wenchang/blob/main/docs/releasing.md). Versions follow SemVer over the public surface: the modules, functions, classes, arguments, tool names and schemas, error payloads, and stored format documented on this site. - While the major version is 0, a breaking change bumps the minor version, and any other change bumps the patch version. - A change to the shipped prompt wording is at least a minor release, because it can change what agents store. - Versions are plain `X.Y.Z`; there are no pre-release tags. - A tag is never moved once its GitHub Release exists; a bad release is fixed by the next patch version. ## Known limitations - **A library, not a server.** There is no CLI, console script, or ready-made server binary. You supply the identity resolver, scope policy, prompt slots, and a short server script or tool registration. The only server in the repository is the dev-only eval server (`python -m evals.server`), which uses in-memory storage. - **One production backend.** `GcsStorage` is the only backend meant for production. `FilesystemStorage` (a local directory, POSIX only) and `InMemoryStorage` are for tests and local development. - **No remote transport.** The only `TransportClient` is `InProcessClient`, which calls a `MemoryStore` in the same process. Tools and storage access always run in your process. - **No content search.** Files are found through `description` and `aliases`. The quality of that metadata is what makes recall work. - **Not snapshots.** `list_prefix` and `get_memory_index` are not point-in-time snapshots: a file written or deleted while an index is being built may or may not appear. `export` has the same property. - **Reads are never role-restricted.** A caller can read any area (including `system/`) of any scope it holds a grant for. `ScopePolicy` gates writes only. - **Privacy is prompt-level.** The prompt tells the agent to refuse financial account numbers, health diagnoses, and anything indicating the user is a minor, and to be stricter in shared scopes. No tool checks content, so a successful write does not mean the content was safe. - **Behavior depends on the agent.** Where facts get filed, and how well, is up to the model following the prompt. The repository's behavioral evals run scripted scenarios against a real agent and grade what it stored; they are signal for prompt iteration, not release gates. - **Concurrency is the caller's concern.** MCP tool calls run concurrently on worker threads and share one client. One `GcsStorage` and its `storage.Client` are safe to share across threads; see [Threads](integrate-without-mcp.md#threads) for caveats. Direct calls are synchronous and block on storage. - **The file cap is soft.** Concurrent creates in one entity can pass `max_files_per_entity`; see [Files per entity](storage.md#file-cap). - **No timeout or retry settings.** `GcsStorage` uses google-cloud-storage's defaults. See [Timeouts and retries](storage.md#retries). --- HTML version: