Agent Skills vs MCP is not really a fight — but on cost, Skills win outright. Anthropic’s own numbers show 58 MCP tools burning roughly 55,000 tokens before an agent does anything useful. A Skill’s metadata costs about 100 tokens. Build Skills for procedure and judgment, MCP servers for live system access, and cache aggressively if you run both.
On August 19, 2026, Anthropic moved Agent Skills, the Skills API, computer use, browser use and the Files API out of beta and into general availability on the Claude Developer Platform, according to the platform’s release notes. Three weeks earlier, the Model Context Protocol shipped its stateless 2026-07-28 specification.
Both are now production infrastructure. Both are free to adopt. Only one of them charges you rent on every single request.
What actually changed on August 19, 2026?
Agent Skills stopped being an experiment. The skills-2025-10-02 beta header is gone, the /v1/skills endpoint is GA, and computer use shipped as computer_toolset_20260801 with browser use as a separate tool, browser_toolset_20260801. That is a full production agent stack in one release.
The Files API went GA the same day. Upload, download, list, metadata and delete operations are free — you are billed only for file content that actually enters a Messages request, at standard input-token rates.
Storage caps are generous: 500 MB per file and 1 TB per organization, rate-limited to roughly 500 requests per minute.
None of the superseded beta headers have a published sunset date yet. Legacy identifiers still work. But GA is the signal that pricing and architecture decisions made today will stick.
What is the difference between Agent Skills and MCP?
MCP gives an agent reach. Agent Skills give an agent method. An MCP server tells Claude how to connect to GitHub and what it can do there. A Skill tells Claude how your team actually writes a release note, in what order, with which checks.
That distinction sounds academic until you look at where each one lives in the context window.
How MCP loads tools
MCP tool definitions are pushed into the model’s context up front, on every request. The agent has to know a tool exists before it can call it, so the full schema — names, parameters, descriptions — sits in the prompt whether or not the tool ever gets used.
That is a fixed tax. It scales linearly with how many servers you connect.
How Agent Skills load
Skills use three-tier progressive disclosure, documented in Anthropic’s Agent Skills overview. Level 1 is YAML frontmatter — name and description only — at roughly 100 tokens per Skill, always loaded. Level 2 is the SKILL.md body, under about 5,000 tokens, loaded only when the Skill is triggered.
Level 3 is where it gets interesting. Bundled reference files and scripts cost zero tokens until read. Script code never enters the context window at all — only its output does.
You can ship 300 pages of API documentation inside a Skill and pay nothing for it unless the agent opens the file.
How much do MCP tool definitions actually cost?
More than most teams realize. Anthropic published a concrete five-server example in its advanced tool use write-up: GitHub at 35 tools (~26K tokens), Slack at 11 tools (~21K), Sentry at 5 (~3K), Grafana at 5 (~3K), and Splunk at 2 (~2K).
That is 58 tools consuming approximately 55,000 tokens before the conversation even starts.
Anthropic’s own internal setup is worse: tool definitions there consume 134,000 tokens before optimization. On a 200K context window, that is two-thirds of the budget spent on a menu the agent mostly ignores.
Now price it. Claude Sonnet 5 costs $2 per million input tokens and Claude Opus 5 costs $5, per the official pricing page. Sonnet 5’s introductory rate was made permanent on August 10, 2026, and the scheduled September increase was cancelled.
What does that cost per month at real volume?
| Setup | Tokens per request | Sonnet 5 cost / request | 10,000 runs / month |
|---|---|---|---|
| Anthropic’s internal tool set | 134,000 | $0.268 | $2,680 |
| 58 MCP tools (5 servers) | ~55,000 | $0.110 | $1,100 |
| Same tools + Tool Search Tool | ~8,700 | $0.017 | $174 |
| 20 Skills + 1 triggered SKILL.md | ~7,000 | $0.014 | $140 |
| 20 Agent Skills (metadata only) | ~2,000 | $0.004 | $40 |
The spread between the top and bottom row is 27x. On Opus 5 at $5 per million input tokens, the same gap costs $6,700 versus $100 a month.
The prompt caching escape hatch
MCP defenders have a real counterargument: cache the tool definitions. Anthropic’s pricing page lists cache reads at 0.1x the base input rate — $0.20 per million tokens on Sonnet 5.
Cached, that 55,000-token block drops from $0.110 to about $0.011 per request. A 1-hour cache write costs 2x base and pays for itself after two reads.
So caching closes most of the gap — if your traffic is dense enough to keep the cache warm and your tool list is stable. Bursty, low-volume agents get cache misses and pay full freight.
Does trimming tools hurt accuracy?
No — it helps. This is the part that surprises people. Anthropic’s Tool Search Tool delivers an 85% reduction in token usage while keeping the full tool library reachable, cutting roughly 77K tokens of overhead down to about 8.7K and preserving 95% of the context window.
Accuracy went up. Opus 4 improved from 49% to 74% on the tool-use benchmark. Opus 4.5 went from 79.5% to 88.1%.
Programmatic Tool Calling shows the same pattern: average usage dropped from 43,588 to 27,297 tokens, a 37% reduction on complex research tasks, while internal knowledge retrieval rose from 25.6% to 28.5% and GIA scores went from 46.5% to 51.2%.
Fewer tools in context means less for the model to confuse. Context bloat is an accuracy problem wearing a cost problem’s clothes.
When should you still build an MCP server?
When you need a live connection, real authentication, or one integration shared across many agents. Skills are static files — they cannot hold an OAuth token, stream an update, or talk to your database. MCP is the transport layer, and nothing about Skills replaces it.
MCP’s adoption numbers back that up. Claude’s connector directory now lists over 950 MCP servers, and MCP has passed 400 million monthly SDK downloads — a 4x increase this year, per Anthropic’s spec announcement. The TypeScript and Python SDKs have crossed a billion total downloads between them.
Use this five-question test before you write a line of either:
- Does it need a live connection to a running system? MCP server.
- Does it need per-user auth or scoped permissions? MCP server.
- Is the hard part procedure and judgment, not access? Agent Skill.
- Is it large reference material used occasionally? Agent Skill — Level 3 files cost nothing until opened.
- Do many agents share one integration? MCP server centralizes it; a Skill travels with the agent.
Agent Skills vs MCP: which should you use for your task?
| What you’re building | Pick | Reason |
|---|---|---|
| Read/write live Slack, GitHub or Jira data | MCP server | Needs a live, authenticated connection |
| House style guide, review checklist, report format | Agent Skill | Pure procedure; ~100 tokens idle |
| Ship 300 pages of API reference to the agent | Agent Skill | Level 3 files cost 0 tokens until read |
| Per-user OAuth scopes and audit trails | MCP server | Auth belongs at the connection layer |
| Deterministic script the agent runs but shouldn’t read | Agent Skill | Script code never enters context |
| One integration consumed by a dozen agents | MCP server | Central updates, single surface |
| Pull live data and apply a fixed workflow | Both | MCP for access, Skill for method |
| Low-volume, bursty agent on a tight budget | Agent Skill | Cold caches make MCP overhead expensive |
What did the 2026-07-28 MCP spec change?
It made MCP stateless, which is the single biggest cost change on the server side. The initialize/initialized handshake and the Mcp-Session-Id header are retired. Every request is now self-describing, so any request can land on any instance behind a plain round-robin load balancer.
That kills the shared-storage requirement. You can run MCP serverless and stop paying for sticky sessions.
Method and tool names now travel in HTTP headers rather than JSON bodies, letting gateways route without parsing payloads. Multi Round-Trip Requests replace server-initiated streams for interactive confirmations.
Roots, Sampling and Logging are deprecated but will keep working for at least twelve months, as will the legacy HTTP+SSE transport. You have a year to migrate — not a weekend.
One number worth sitting with: at Honeycomb, nearly 20% of all monthly interactive queries are now made by agents, not humans. That ratio is why the per-request tax matters.
Frequently asked questions
Are Agent Skills a replacement for MCP?
No. Skills carry procedural knowledge as files; MCP carries live, authenticated access. Anthropic shipped both to GA in 2026 and the standard production pattern uses them together.
How many tokens does one Agent Skill cost?
Roughly 100 tokens for its metadata at startup, per Anthropic’s documentation. The SKILL.md body — under about 5,000 tokens — loads only when the Skill is triggered.
Do MCP tool definitions get charged on every request?
Yes, unless cached. Anthropic’s example of 58 tools across five servers consumes roughly 55,000 input tokens per request. Prompt caching cuts that to 0.1x the base rate on cache hits.
Does the Files API cost extra?
No. Upload, download, list, metadata and delete are free. You pay only for file content that enters a Messages request, billed as normal input tokens.
Is the old MCP spec still supported?
Yes. Roots, Sampling, Logging and the HTTP+SSE transport are deprecated but supported for at least twelve months from the 2026-07-28 release.
Which model should I run agents on to keep costs down?
Claude Sonnet 5 at $2/$10 per million tokens is the volume workhorse; its introductory pricing was made permanent on August 10, 2026. Opus 5 at $5/$25 is 2.5x the input cost — worth it only when the task genuinely needs it.
Do fewer tools in context make agents dumber?
The opposite. With Tool Search Tool enabled, Opus 4.5 improved from 79.5% to 88.1% on Anthropic’s tool-use benchmark while using 85% fewer tokens.
The bottom line
Build Skills first. They are nearly free to keep loaded, they went GA on August 19, and Anthropic’s own benchmarks show that trimming context raises accuracy rather than lowering it.
Add MCP servers only where you need a live connection, real auth, or one integration shared across many agents — then wrap them in Tool Search Tool and prompt caching on day one, not after the first surprising invoice.
The decision rule is exactly this: if the hard part is reaching the system, build MCP; if the hard part is knowing what to do once you’re there, build a Skill. Teams running high-volume agents on the naive pattern are paying somewhere between 8x and 27x more per request than they need to, and getting worse answers for the money.
Related reading on agent economics: our breakdown of Claude Code vs Codex CLI on cost, the Gemini 3.7 Flash vs Claude Sonnet 5 cost-per-coding-point comparison, and ChatGPT Business vs Claude Team on the seat-license side.
Sources
- Model Context Protocol — The 2026-07-28 Specification
- Anthropic — MCP 2026-07-28 spec: stateless core, coming to Claude
- Anthropic Engineering — Advanced tool use on the Claude Developer Platform
- Claude Platform Docs — Agent Skills overview
- Claude Platform Docs — Pricing
- Claude Developer Platform release notes, August 2026
Leave a Reply