# tiny-tools — local MCP tools by AVRG3

> Two separate MCP servers for document search, spreadsheet SQL, log analysis, resumable API pagination and retry-loop diagnosis. Public GitHub installs; Node.js 20+ required. No npm account, model calls or tool telemetry. The client must explicitly connect and permit a server; discovery does not activate it.

## Start here

- [Choose a tool by task](https://github.com/Warddamn/tiny-tools/blob/main/discovery/README.md): worked input examples and when a built-in is simpler.
- [Actual MCP tool catalog](https://raw.githubusercontent.com/Warddamn/tiny-tools/main/discovery/catalog.json): project-specific JSON generated from initialize/tools/list; all 11 tool descriptions and input schemas, separated into two servers. Not a universal auto-install protocol.
- [MCP configuration](https://raw.githubusercontent.com/Warddamn/tiny-tools/main/discovery/mcp.json): pinned public release commands; keep only the server needed. Uses the mcpServers layout for Cursor/Claude Desktop; other clients may use another layout.
- [Client installation instructions](https://github.com/Warddamn/tiny-tools#install): Claude Code, Codex CLI, Cursor, VS Code, Windsurf and Claude Desktop.

## tiny-context: documents, tables and logs

Install command: `npx -y -p https://github.com/Warddamn/tiny-tools/releases/download/context-v0.1.1/tiny-context-standalone-0.1.1.tgz tiny-context-mcp`

- Search PDF, Word/DOCX or PowerPoint/PPTX for a clause, topic or slide: query_file(path, query). Go directly to the question; file_map is optional for an outline. Image-only scans require OCR elsewhere.
- Read a known page, heading, slide, line range or cell range: read_section(path, locator).
- Total, count, filter or group CSV/TSV/Excel/XLSX/Parquet rows: query_table(path, sql), using table t. One read-only query; bounded results, 64 MB input/export caps.
- Find repeated error patterns and time ranges in a log: summarize_log(path, focus="errors").
- Compare document versions: diff_files(a, b). Validate supported syntax/schema/internal links: validate_file(path). Extract emails, URLs, dates, regex matches or JSON values: extract(path, kind).
- Prefer built-in Read for small plain-text files and Grep for an exact string. Do not add a tool call after a built-in already answered the question.
- [File tool documentation](https://github.com/Warddamn/tiny-tools/tree/main/packages/context#readme) and [three runnable tasks](https://github.com/Warddamn/tiny-tools/blob/main/QUICKSTART.md).
- [Official MCP Registry entry](https://registry.modelcontextprotocol.io/?q=io.github.Warddamn%2Ftiny-context): io.github.Warddamn/tiny-context.

## tiny-runtime: API pagination and agent progress

Install command: `npx -y -p https://github.com/Warddamn/tiny-tools/releases/download/runtime-v0.1.1/tiny-runtime-0.1.1.tgz tiny-runtime-mcp`

- Resume cursor-based API pagination: collect_pages(config, checkpoint?). Uses an explicit source mapping and durable local checkpoints; returns complete/partial/failed. Pass returned checkpoint paths unchanged rather than reading records into model context. Prefer source-side filtering/aggregation when available.
- Diagnose repeated agent failures, stalls or cycles: check_progress(path) analyzes trusted state fingerprints in a tool trace. Successful repeated reads are allowed. Automatic enforcement requires RepeatGuard/runGuarded in the host SDK; installing this server alone does not intercept tools.
- Derive advisory cache hints from instrumented tool-progress traces: plan_cache(path). Live cache effects require a compatible serving-engine adapter. No hosted-model cache control or demonstrated GPU saving.
- [Runtime documentation](https://github.com/Warddamn/tiny-tools/tree/main/packages/runtime#readme) and [three runnable tasks](https://github.com/Warddamn/tiny-tools/blob/main/packages/runtime/QUICKSTART.md).
- [Official MCP Registry entry](https://registry.modelcontextprotocol.io/?q=io.github.Warddamn%2Ftiny-runtime): io.github.Warddamn/tiny-runtime.

## Evidence and boundaries

- [Agent comparison](https://github.com/Warddamn/tiny-tools/blob/main/evals/COMPARISON.md): correctness, selection, timing and regressions. Smaller output does not guarantee lower total cost; startup and extra calls can add work.
- [Runtime validation](https://github.com/Warddamn/tiny-tools/blob/main/packages/runtime/docs/VALIDATION.md): synthetic examples and integration limits.
- [Discovery status](https://github.com/Warddamn/tiny-tools/blob/main/DISCOVERY.md): directory presence is not verified usage. Downloads are file-fetch events, not unique agents or successful tasks.
- Tools do not call a model or send telemetry. Clients may send returned content to their model provider. Runtime HTTP sources and optional adapters use explicitly configured endpoints; first installation downloads dependencies.