Capo: Community-Driven AWS SDK for Python
capo-sdk.dev- Category
- Developer Tools
- Rank
- No. 1367Tools index
Previous survey · No. 1322 ·
- Type
- TOOL
- Use case
- Deployment & Operations
- GitHub
- 70 stars
- Latest release
- release-2026-11-06
- Date
About
Capo is a community-driven AWS SDK for Python in which every AWS service is a standalone package with its own documentation site. Each site covers that service's clients, operations, paginators, waiters, types and errors. The docs also cover streaming, interceptors, HTTP caching and Pyodide (WASM) support.
At a glance
Capo is a community Python SDK for AWS. It is worth a small trial if you want a typed asynchronous client and only need packages for the services you use. The generated clients derive their request and response types from AWS service models.
Good fit
Python developers adding an AWS integration who want type checking and synchronous or asynchronous calls.
Consider something else when
Skip it if you require AWS’s official SDK or need to keep an existing Boto3 integration unchanged.
Start here
Pick one read-only operation your service already uses. Install the matching service package in a scratch project, check its types, and compare the response with your existing client using a restricted test account.
Cost and requirements
MIT-licensed. The S3 package requires Python 3.10 or later. Calls still need AWS credentials, permissions and a region, and AWS service charges can apply.
Limits to know
The maintainers describe the SDK as early in development and advise pinning versions. An unknown union variant can be surfaced by name without its payload, so handle new response variants explicitly.
Related work
Sources checked 2026-11-05. Source-based assessment.
Read the source review
What it does
A Python client library for Amazon Web Services, generated from the Smithy API models Amazon publishes, where each service installs as its own small package. Every service gets an async and a sync client with identical methods. Requests take keyword arguments and plain nested dictionaries that a type checker still verifies, and responses come back in the same typed shapes. Pagination, waiters, presigned URLs, the standard credential chain and an HTTP/2 transport built on the zapros client are included, and the README states it runs in the browser through Pyodide.
Why it's ranked here
The case rests on a gap between two incumbents, and the project's own comparison argues it concretely. By its account boto3 is sync only and untyped without third-party stubs, while Amazon's experimental successor is async only, publishes 25 clients and has no paginators, waiters or presigned URLs. Capo claims every service in the Smithy models, both calling styles, asyncio and trio, and a Bedrock Runtime install of 912 KiB against 16,485 KiB for boto3. Those figures are the author's, measured on one machine, but the commands to reproduce them are printed. The counterweight is maturity: the README warns of breaking changes before the first major release.
What's good
Inputs and outputs share one set of typed dictionary definitions, so a response can feed straight into another call, and nested inputs need no extra imports. Configuration resolves in a stated order: per-call override, client argument, environment variable, then profile. Pointing one service at a local S3-compatible server such as MinIO or LocalStack is a one-line change. Credentials follow the familiar chain, including IAM Identity Center, web identity for EKS and container endpoints, with a caching wrapper that refreshes 60 seconds before expiry. The integration test setup separates tests that cost money on AWS and skips them unless explicitly enabled.
Tradeoffs
The README calls the SDK early development and recommends pinning a version, so breaking changes are expected. AssumeRole, web identity and IAM Identity Center credentials need an optional extra installed on each service package. The head-to-head comparison is written by the project and dated, so its claims about the official successor will age. Coverage of more than 440 services comes from code generation, while the integration test project installs only nine service packages, so most services rest on the generator being right. One package per service also means more dependencies to keep in step.
How to use it well
Fits a typed Python codebase, especially async code on asyncio or trio, that calls a handful of AWS services and wants editor completion without stub packages. Install only the service packages you use, pin the version, and add the single sign-on extra if you rely on role or Identity Center profiles. Credentials resolve on every request, so wrap any provider that makes network calls in the caching provider. For local work, override the endpoint to an S3-compatible server. Large boto3 codebases face a rewrite rather than a swap: argument naming and nested shapes differ.
Technical notes+
README.md documents configuration precedence (config_overrides, then client argument, then environment variable, then profile) and a seven-step default credential chain, where the role, web identity and SSO steps require the sso extra; CachedProvider re-resolves 60 seconds before expiration. docs/index.md, docs/usage.md, docs/configuration.md, docs/authentication.md, docs/forward-compatibility.md and docs/wasm.md are each a single snippet include of a README section. The root pyproject.toml is an unpublished test harness (requires-python >=3.10, uv package = false) that installs nine capo service packages as editable path sources under services/, pulls pytest, anyio, trio, pyodide-build and zapros with the caching extra, and defines pytest markers third_party_services (local S3-compatible servers via docker compose) and paid (skipped unless CAPO_PAID=1). package.json is a private npm workspace over services/* that uses Changesets for versioning, followed by two Python sync and changelog helpers. docs/comparison.md attributes boto3's install size to botocore shipping every service model and shows capo-bedrock-runtime depending only on zapros and small pure-Python packages.
Observed
- Language
- Python 3.10 or later
- Packaging
- One installable package per AWS service, published under the capo- prefix
- Interface
- Python library with matching async and sync clients per service
- Async runtimes
- asyncio and trio
- Typing
- Inputs and outputs typed as TypedDicts, interchangeable between calls
- Code generation
- Generated from the official AWS Smithy models
- HTTP layer
- Built on the zapros HTTP client, with HTTP/2 support
- Platform support
- README states it runs in WASM environments via Pyodide
- Stability
- README describes early development and recommends pinning the version
- Tests
- Integration tests live in an unpublished root project; tests that cost money on AWS are opt-in
Read from README.md, pyproject.toml, package.json, docs/index.md, docs/authentication.md, docs/comparison.md, docs/configuration.md, docs/usage.md, docs/forward-compatibility.md, docs/wasm.md.
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.