Changelog

2.0.1 - 2026-11-08

• Add an explicit Claude manifest for claude.ai marketplace synchronization.

2.0.0 - 2026-11-08

• Remove the bundled personal MCP endpoint. Configure an HTTP server named firecrawl in your client before upgrading; see the README.
• Keep both skills and their names. Point the Claude research agent at the user-configured server and remove machine-specific operator notes.

1.2.0 - 2026-10-27

• firecrawl-research answers people in plain prose with cited sources. The Answer / Confidence / Status block moved to references/answer-format.md and is used only when a prompt or harness asks for it, such as eval runs.
• Sharper skill description naming the built-in WebSearch and WebFetch, so ordinary current-information questions load the skill.
• README: a ~/.claude/CLAUDE.md snippet that makes Claude Code use the skill for web questions. In headless Sonnet runs the description alone loaded the skill 0/3 times; with the snippet, 3/3.

1.1.0 - 2026-10-27

• Added the Claude Code personal-firecrawl:web-researcher agent (claude/agents/web-researcher.md, wired through the Claude marketplace entry). Its tools allowlist is only the plugin's Firecrawl search and fetch, so Claude Code enforces the limit and it cannot start subagents.
• firecrawl-research uses that agent on Claude Code and keeps the web-researcher brief for other clients. Independent attempts paste the puzzle method into the agent's prompt.
• The brief now checks status, error and truncated and retries failed fetches through the Wayback Machine, matching the skill.
• Noted that the hosted search route returns at most 10 results.

1.0.1 - 2026-10-26

• firecrawl-research: Ceramic-aware query rules (plain keywords before "exact phrase", OR and exact-host site:; default 10 results), a structured-sources reference for papers, archives and lists, readings for vague phrases, a Wayback retry for failed fetches, and independent attempts with reconciliation for hard unverified answers.

1.0.0 - 2026-10-26

Breaking: seven skills removed and replaced by two.

• Added firecrawl-research: all web lookups and research through the MCP search and fetch tools, with question sizing, keyword query rules, a multi-clue puzzle method, reading rules, subagent fan-out, and a fixed answer format (best answer, confidence, verified or unverified, sources actually seen).
• Rebuilt firecrawl-cli for on-host bulk work only: map, crawl, download and saved files through the local CLI. Kept the endpoint, CLI and security references.
• Retired firecrawl-search, firecrawl-scrape, firecrawl-map, firecrawl-crawl, firecrawl-download, firecrawl-agent and firecrawl-interact. Server-side deep research is no longer described; the wrapper no longer exposes it.
• Shipped the web-researcher subagent as a brief inside firecrawl-research rather than an agent file, because Agent Plugins agents are client-specific.
• MCP connection unchanged.

0.2.0 - 2026-10-21

• Migrated MCP instructions to search, fetch and deep_research for wrapper 3.0.
• Updated inputs, bounded structured results and research ID continuation.
• Routed removed map, crawl, batch, extract, agent and action operations to supported on-host CLI/API paths without promising unavailable MCP tools.
• Preserved all eight skill names and the production OAuth connection.

0.1.0 - 2026-08-27

• Grouped eight self-hosted Firecrawl skills into one marketplace plugin.
• Added the authenticated production Firecrawl MCP endpoint without credentials.
• Updated the skills for the personal wrapper's scrape actions, structured extraction, asynchronous jobs, agent workflow, and deep-research tools.
• Kept local CLI fallbacks for file output and options not exposed by the MCP wrapper.