Firecrawl turns private AI sessions into shared team work with Claude Tag

AI work stuck in one person’s session

Before Claude Tag, Firecrawl’s engineers worked in Claude Code, teammates ran their own Claude chats, and in January the company built a Claude-powered bot for its support system. Most of that work happened in sessions only one person could see, so the weak point was the handoff. “Someone could investigate a problem with AI, but the rest of the team saw only the answer or the pull request they chose to share,” said Eric Ciarla, cofounder and Chief Growth Officer at Firecrawl. Whoever picked up the work next had to rebuild the background, assumptions and failed attempts still sitting in the first person’s session.

Support felt the cost most, because one customer question can need code knowledge, account context and product history, often pulling several people in. Each time, an engineer had to stop and reconstruct the problem before work on an answer could start, and support’s own requests for better internal tools added one more handoff before a customer got an answer.

Claude where the team already talks

In late June, Firecrawl began moving work from its own bot to Claude Tag one workflow at a time. Claude Tag reads the whole Slack thread and posts its work back into it, where the team can see and steer it, which Firecrawl’s bot couldn’t do. Firecrawl also doesn’t have to maintain the sandboxes Claude Tag runs in, and it now keeps its own bot only for jobs wired into its support system. Firecrawl moved Claude Tag from Claude Opus 5 to Claude Opus 5.5, with its 1M-token context window, within two days of its September release, because it is faster, does better work and costs less to run.

The rollout began in the support and engineering channels and in the channel for Firebrain, Firecrawl’s internal collection of company knowledge, skills and tools. The team then connected GitHub and read-only BigQuery analytics. Each channel carries its own shared instructions and tool access, so a new teammate can tag Claude with no local setup.

“Claude posts its investigations and progress as an agent in the thread, which helps people see what it did and what still needs review,” Ciarla explained.

One thread from screenshot to merge

On September 15, Ciarla posted a screenshot in #ai-engineers showing divider lines spilling into Firecrawl’s navigation and tagged @Claude. Claude traced the problem to four settings cards and had a pull request up 9 minutes 30 seconds later with 127 unit tests passing. Because the live page sat behind a login, it reported that it had checked the layout on an isolated copy, so Ciarla knew what was still unverified. He reviewed and merged the fix less than 18 minutes after posting the screenshot.

The thread works for requests from outside engineering too. Micah Stairs, Head of Support Engineering, asked Claude for a way to look up teams by their slug, a short text identifier, in his team’s admin panel, and Gaurav Chadha, a Developer Support Engineer, tested the pull request. Chadha found that generated default slugs needed to work too, and Claude updated the same pull request. The change merged about three hours after the first request, all in one conversation.

The thread also keeps wrong turns visible. In Firecrawl’s agent product, a follow-up question was showing the previous answer’s results table. Claude first blamed how the frontend polled for results, while flagging the server as a possibility. When Claude tested its own fix, the bug was still there, so it kept looking and found the real cause in the agent’s backend sandbox, which saved files one folder too deep. Its nine-line fix passed all four production QA checks, and the team closed the frontend pull request instead of shipping a fix that wouldn’t have worked.