Your AI coding agent just got better at Fastly

If you've used an AI coding agent to configure a Fastly service, you've probably seen it struggle. Maybe it hallucinated an API endpoint that doesn't exist. Maybe it spent a dozen turns probing the CLI for the right flags. Maybe it wrote VCL that compiled fine, but silently broke grace detection. AI models are trained on past data and Fastly-specific patterns, the kind that make the difference between code that works and code that works well, often aren't in that training data.

We've been thinking about this problem. Earlier this year, we released the Fastly MCP Server to give AI agents access to Fastly through the CLI. That solved the "how do I talk to Fastly?" problem. But there's a second problem: even with tool access, agents don't always know what to do.

Today, we're releasing the Fastly Agent Toolkit to fix that.

Skills, not just tools

The Agent Toolkit is an open source collection of AI agent skills. These are curated knowledge files that teach your coding agent how to work with the Fastly platform. Think of it as giving your agent the experience of a senior Fastly developer.

The toolkit includes six skills:

1. fastly - Platform concepts: services, caching, VCL, WAF, TLS, DDoS protection, purging, and API usage

2. fastly-cli - CLI workflows: managing services, deploying Compute apps, configuring logging, working with KV stores

3. falco - VCL development: linting, testing, simulation, formatting, and Terraform integration

4. xvcl - The XVCL transpiler: syntax extensions, subroutines, header manipulation, caching logic

5. viceroy - Local Compute testing with the Wasm-based runner

6. fastlike - Local Compute testing with the Go-based runner

Each skill is a set of markdown files with a SKILL.md entrypoint and a references/ directory containing detailed topic files. Skills are loaded on demand. Your agent only pulls in what's relevant to the current task, so there's no wasted context.

The numbers speak for themselves

We didn't just build skills and hope they worked. In our internal testing across 7 models on real-world Fastly tasks, we consistently saw measurable improvements.

The results:

• 1.6x average wall-clock speedup

• 30-70% fewer tool calls

• Skills won 7 out of 8 tasks head-to-head against baseline

Here's what that looks like in practice. We asked an agent to configure a Fastly service to cache content from a specific domain and deploy it. Without skills, the agent spent 25 turns over 11 minutes, reading JSON config files, trying wrong flag combinations, and manually managing versions. With skills, it loaded the fastly-cli skill on the first turn, found the right commands, used --autoclone on the first try, and finished in 9 turns and under 3 minutes. 4.1x faster, 2.8x fewer turns.

The biggest improvements show up where Fastly-specific knowledge matters most: VCL patterns like subfield() for cookie parsing, obj.ttl == 0s for grace detection, correct Vary header handling, and the right deploy sequence for Compute services. These are exactly the patterns that models don't have in their training data and can't figure out by trial and error alone.

Works with your agent

Skills follow the open Agent Skills spec and work with any agent that reads markdown. That's a long and growing list: Claude Code, Gemini CLI, Codex, Cursor, Amp, Cline, Kimi Code, OpenCode, Replit Agent, Warp, and more.

Get started in 60 seconds

Install the skills you need with the skills CLI:

# Install into your project
npx skills add github:fastly/fastly-agent-toolkit --skill fastly-cli --skill viceroy

# Or install globally
npx skills add -g github:fastly/fastly-agent-toolkit --skill fastly-cli --skill viceroy

If you use Claude Code, you can also install from the plugin marketplace:

claude plugin install fastly-agent-toolkit@claude-plugins-official

Or just copy the skill directories manually. They're plain markdown files:

mkdir -p .agents/skills
cp -R skills/{fastly-cli,viceroy} .agents/skills/

Pick the skills relevant to your work. For VCL, grab falco and xvcl. For Compute, grab fastly-cli and either viceroy or fastlike. The fastly skill covers the platform broadly and pairs well with anything.

MCP + Skills: better together

If you're already using the Fastly MCP Server, skills are the perfect complement. MCP gives your agent the ability to act - creating services, purging caches, and deploying code. Skills give it the knowledge to act correctly. Knowing the right patterns, the right flags, and the right order of operations. Together, they represent our commitment to giving you the best, most modern developer experience on any edge platform.

Open source, open to contributions

The toolkit is open source on GitHub. Skills are just markdown. You can read them, modify them, and contribute your own. If you've found a pattern that makes your agent better at Fastly, we want to hear about it.

We'd love your feedback. Try the skills, see how they change the way your agent works with Fastly, and let us know. Tell us what worked well, what didn't, and what patterns or workflows you'd like to see added. Share your results in the Fastly Community Forum.

We're committed to continuously improving these skills and the developer experience around them - measured, tested, and driven by how you actually use Fastly. Using AI agents with Fastly just got a lot better.