~alter mcp-ollama

Extract structured information from text locally. Use this for classification, entity extraction, tagging, pattern recognition, or any task that transforms unstructured text into structured output. Good for: categorising issues, extracting names/dates/amounts, tagging content, parsing semi-structured data, sentiment analysis.

Code-aware local generation. Claude reads source code with Read/Grep (free), then delegates the TEXT GENERATION to local - docstrings, test stubs, explanations, type annotations, inline comments, or review feedback. This is the primary tool for reducing API token usage on code tasks. Claude orchestrates (decides what code to read, what task to perform), but the actual generation happens locally at zero API cost. Accepts up to ~12K tokens of code context (16K model context minus overhead). For larger contexts, break into focused chunks (one function, one class).

Analyse a git diff locally. Use this for generating commit messages, PR descriptions, changelog entries, or understanding what changed in a diff. Claude runs `git diff` (free), passes the output here for text generation. Particularly valuable for large diffs that would consume many API tokens if processed by Opus/Sonnet directly.

Draft text locally - commit messages, PR descriptions, docstrings, changelog entries, documentation sections, or any formulaic text that follows a template or convention. Saves Opus tokens on boilerplate generation.

General-purpose local LLM generation. Delegates work to a local Ollama model to save tokens. Use this when the task is bulk text processing that doesn't require Opus-level reasoning: summarising documents, explaining code for docs, generating boilerplate, reformatting content, translating, etc. The local model (default: hermes3:8b) runs on the user's hardware - zero API cost, ~40 tok/s on a decent GPU.

List all models available in the local Ollama instance. Use this to check what models are loaded and available for delegation.

Pull/download a model into Ollama from the registry, or import a local GGUF file. Use this to make additional models available for delegation. For GGUF files, create an Ollama Modelfile first, then use 'ollama create'. This tool handles registry pulls (e.g., 'qwen2.5:14b', 'deepseek-r1:8b').

Summarise long text locally. Use this instead of having Opus process large blocks of text when you only need a summary. Feed in file contents, docs, logs, or any bulk text and get a concise summary back. Particularly valuable for: large file contents, documentation, log output, meeting notes, long git diffs, error traces.

Perform mechanical code transformations locally. These are pattern-based transforms that don't require architectural understanding - the kind of work that burns API tokens for no good reason. Use this for: converting between formats, renaming patterns, migrating syntax, generating boilerplate from examples, etc.

Have a local vision model SEE a rendered image (screenshot of a UI) and report what is literally on screen - for zero API cost. local_analyze is TEXT-ONLY; this is the multimodal path. It REPORTS (is the main content populated or blank/error, what regions exist, any illegible/clipped text, what the eye lands on first). It does NOT rank severity, judge on-brand, or greenlight a change - a local vision model reads pixels well but ranks badly, so keep the judgment on the calling model. Modes: 'see' (full structural report), 'emptystate' (populated-vs-degraded data check), 'legibility' (readability faults). Feed near-full-resolution PNGs; downscaling below ~1280px wide makes the model hallucinate data presence.