Kimi K3 + OpenCode + Firecrawl MCP Web-Connected Agent Workflow

One-sentence takeaway

This complete case study connects Kimi K3 to OpenCode and Firecrawl MCP, allowing it to call live web search during long-running coding and research tasks.

Use cases

• Suitable tasks: Tasks that require current web sources, long-context codebases, browser automation, or research-oriented agents.

• Unsuitable tasks: Short answers that do not need web access, and environments with hard requirements for a stable native Windows terminal where WSL cannot be used.

• Applicable model version: moonshotai/kimi-k3; the article also says that the weights were released on 2026-07-27.

• Applicable client, agent, or API: OpenCode, Firecrawl CLI/MCP, OpenRouter, or another Kimi K3 provider.

• Recommended reasoning tier and parameters: The article does not provide fixed parameters; for long-running agent tasks, use K3’s default thinking mode and let the harness manage the tool loop.

Ready-to-use content

Installation and launch:

npm install -g opencode-ai
opencode auth login
opencode

npm install -g firecrawl-cli
firecrawl setup mcp
firecrawl launch opencode

Verify MCP in OpenCode:

do a search for the latest news using the firecrawl mcp

Test/workflow steps

1. Install OpenCode and log in to the selected model provider, confirming that the model name is moonshotai/kimi-k3.

2. Install the Firecrawl CLI; run firecrawl setup mcp to write Firecrawl MCP into the OpenCode configuration.

3. If the CLI refuses to write the API key to the client configuration after firecrawl login has been run, export FIRECRAWL_API_KEY instead, or launch with firecrawl launch opencode.

4. Restart OpenCode so Kimi K3 can discover the MCP tools.

5. Use the test request above to trigger a search; check that the model actually calls Firecrawl instead of answering from memory.

6. For research tasks, add output constraints requiring sources and dates and requiring uncertainty to be stated explicitly, then save the search results and final deliverable.

Original evidence and data

• The article explicitly provides commands for OpenCode, the Firecrawl CLI, setup mcp, and launch opencode, along with one web-search request.

• The article says K3 can obtain live web data through Firecrawl MCP; it does not provide an independent comparison of success rate, latency, or cost.

• The native Windows workflow encountered problems, and the author recommends Ubuntu/WSL; this is an observation from the author’s environment, not a Kimi official compatibility commitment.

Scope and limitations

• The article also promotes the Firecrawl service, and its web-access conclusion comes from a single case; it should not be treated as a controlled agent benchmark.

• OpenCode, the Firecrawl CLI, the MCP schema, and provider versions were not fixed; record versions and routing when reproducing it.

• MCP tools expand the data and operations accessible to the model. In production, configure domain, permission, budget, and human-approval boundaries.

Reproduction steps

Run the installation commands above in an isolated project directory and record the CLI/OpenCode versions. Run the same test request three times, saving tool-call logs, retrieved URLs, citations in the answer, total tokens, and the number of failed retries. Compare “without MCP” and “with MCP” groups to determine the actual benefit of the web-access tool.

Source excerpt or observation (brief excerpt for compliance only)

The original article says: “We got Kimi K3 running in OpenCode and gave it live web access using the Firecrawl MCP.”