zakira-recall

name	zakira-recall
description	Search the web, fetch pages, and run multi-source research with citations using Zakira.Recall, a local CLI and MCP server that uses Playwright-driven providers (DuckDuckGo, Bing) without API keys. Use this when the user wants web search results, wants to read or extract content from a URL, wants deep research with sources, when configuring profiles or browser authentication for blocked providers, or when troubleshooting failed searches, consent pages, captchas, or provider fallback. Works both through the `recall` CLI and through the `recall mcp` stdio server (tools `WebSearch`, `WebFetch`, `WebResearch`, `WebBatchFetch`, `WebSearchThenFetch`, `WebListProviders`, `WebGetProviderHealth`, `WebShowConfig`, `WebShowProfile`).
license	MIT
compatibility	Requires the `recall` global tool (`dotnet tool install --global Zakira.Recall`) and a Playwright chromium install. Works on Windows, macOS, and Linux. CLI and MCP stdio server.
metadata	{"project":"Zakira.Recall","homepage":"https://github.com/MoaidHathot/Zakira.Recall","version":"0.5.0"}

Zakira.Recall

Zakira.Recall is a local web search, page fetch, and research toolkit for AI agents. It exposes the same capabilities through two surfaces:

• CLI: the recall global tool. Use this in shell scripts and one-off commands.
• MCP server: recall mcp runs an MCP stdio server that exposes typed tools to MCP-compatible clients (Claude Code, OpenCode, Cursor, etc.).

Providers are Playwright-backed (duckduckgo, duckduckgo-browser, bing) and require no API keys. The tool keeps cookies and consent state in named browser profiles, so a one-time interactive sign-in unblocks providers that show consent pages or captchas.

When to use this skill

Load this skill when the user asks for any of:

• web search ("search for...", "find pages about...", "google this", "look up...")
• reading a URL ("fetch this page", "read https://...", "extract text from...")
• multi-source research with citations ("research...", "what does the web say about...", "summarize sources on...")
• setting up or installing recall, Zakira.Recall, or the MCP server
• diagnosing failed searches, consent prompts, captchas, provider fallback, or health issues

How to navigate this skill

SKILL.md is intentionally small. Load the reference file for the specific task. Each reference is self-contained and < 300 lines so it stays cheap to load.

If the task is...	Load this file	Covers
Search the web for results (titles, URLs, snippets)	references/search.md	WebSearch / recall search, operators, providers, pagination, time range, fallback
Deep research with citations and extracted page content	references/research.md	WebResearch / recall research, top-pages selection, domain diversity, partial errors
Read or extract text from one or more known URLs	references/fetch.md	WebFetch, WebBatchFetch, WebSearchThenFetch / recall fetch
Install, configure, or register the MCP server with a client	references/setup.md	dotnet tool install, Playwright install, recall config init, MCP registration
Searches fail, provider is unhealthy, consent or captcha shown	references/troubleshooting.md	recall profile auth interactive, fallback, health, common errors
Need full MCP tool surface and parameter reference	references/mcp-tools.md	All MCP tools with full parameter tables and return shapes

Decision rules

Pick the right tool the first time:

• One query, one SERP → WebSearch / recall search. See references/search.md.
• One known URL → WebFetch / recall fetch <url>. See references/fetch.md.
• Multiple known URLs → WebBatchFetch. See references/fetch.md.
• Search and then fetch specific result indexes → WebSearchThenFetch. See references/fetch.md.
• "Research this topic" (auto: search + fetch top N + citations) → WebResearch / recall research. See references/research.md.

Use research when the user wants synthesis from multiple sources. Use search when the user wants the result list itself. Use fetch when the user already has the URL.

Output mode (CLI only)

CLI commands default to human-readable text output. When you call the CLI from a script or pipe its output back to an agent, prefer --output json for machine-parseable structured output. Other modes: markdown, dump.

MCP tools always return structured typed objects.

Provider quick reference

• duckduckgo — fastest, no consent in most regions. Default.
• duckduckgo-browser — same provider via Playwright, useful when the HTTP variant is blocked or rate-limited.
• bing — Playwright-driven. May require a one-time consent click via recall profile auth interactive --provider bing.
• Aliases (e.g. ddg) are resolved by the registry, so use them anywhere a provider is accepted.

For provider selection rules and fallback, see references/search.md.

Minimal examples

# Search
recall search "site:github.com mcp server" --limit 10 --output json

# Fetch one URL
recall fetch "https://example.com" --output markdown

# Research a topic with citations
recall research "best local mcp web search tools" --top-pages 3 --output json

# Run the MCP stdio server
recall mcp

For complete examples and parameter tables, load the appropriate reference file from the table above.