Dependency Footprint

textfsm-ai supports 18 LLM providers, but no single install needs all of their SDKs at once. Since v0.6.0, pip install textfsm-ai installs only the core CLI/API — every provider SDK is an opt-in extra (see Installation). This page documents exactly what each extra pulls in, verified with real, clean-venv installs — useful when you're sizing a container image or just curious where the weight goes.

Bare install: 5 packages

pip install textfsm-ai

Installs only textfsm_ai itself plus its 4 core dependencies: PyYAML, requests, click, textfsm (tomli also lists as a dependency, but only actually installs on Python <3.11).

This is fully functional on its own — import textfsm_ai, the CLI, and textfsm-ai providers list all work. Only using a provider requires its extra:

>>> import textfsm_ai
>>> textfsm_ai.generate("sample", provider="anthropic", api_key="x", model="y")
ImportError: Provider 'anthropic' requires additional dependencies that
are not installed. Install with: pip install textfsm-ai[anthropic]

Per-provider package counts

Extra	Total packages	What makes up the difference
[azure]	14	Lightest — reuses requests directly, no httpx/pydantic stack
[bedrock]	16	botocore is the single largest file (~15MB, bundles every AWS service's API definitions), but pulls few extra packages
[oci]	22	Per-request cryptographic signing needs cryptography + pyOpenSSL + PyJWT (no bearer API key at all)
[openai] (+ 9 aliases below)	24	httpx (sync+async client) + pydantic (typed request/response models)
[anthropic]	24	Same httpx + pydantic stack as openai, plus jiter/distro/docstring_parser
[gemini] / [vertexai]	31	httpx+pydantic, plus google-auth → cryptography + pyasn1 (for Application Default Credentials)
[cohere]	31	Unexpectedly pulls in tokenizers + all of huggingface_hub (local token counting)
[mistral]	34	Heaviest — a full OpenTelemetry SDK (opentelemetry-api/-sdk/exporters) plus protobuf, googleapis-common-protos, and invoke (a task-runner library)

[openai] also covers deepseek, groq, xai, together, fireworks, cerebras, perplexity, openrouter, and moonshot at no extra package cost — all nine subclass the same OpenAI-compatible chat- completions client and need nothing beyond the openai package itself.

Package count and disk size are two different axes. oci downloads the single largest file (~36MB) but installs fewer total packages (22) than cohere or mistral (31/34) — a heavy SDK author can produce either a few large files or many small ones.

Why some SDKs are heavier than others

The differences trace back to each provider's own upstream SDK design, not anything this package controls:

• HTTP client choice: httpx (async-capable, used by openai, anthropic, gemini, cohere, mistral) pulls in httpcore, h11, anyio, and sniffio on top of itself. azure and bedrock reuse the plain, sync-only requests this package already depends on core.
• Request/response validation: pydantic (used by most of the httpx-based SDKs) brings its own compiled-Rust pydantic_core plus annotated-types/typing-inspection.
• Credential machinery: gemini/vertexai need google-auth for Application Default Credentials, which needs cryptography to verify signed service-account JWTs; oci needs cryptography+pyOpenSSL+ PyJWT because it signs every request itself rather than sending a bearer token.
• Unrelated SDK features: cohere's tokenizers dependency (for local token counting) and mistral's OpenTelemetry stack (for built-in tracing) have nothing to do with making a chat-completions call, but ship as hard requirements of those SDKs regardless.

How this was verified

Every number above comes from an actual pip install into a fresh, empty virtual environment (not from reading pyproject.toml alone), followed by pip list and a functional check that the installed provider resolves via get_provider_by_name() while every other provider correctly raises the pip install textfsm-ai[...] error. See textfsm_ai/providers/registry.py for the lazy-loading mechanism that makes this possible.