MCP Cost Management: How to Track, Budget & Cut MCP Server and Tool-Call Costs
Will MCP publish a new final specification version by May 2027?
Why the forecast moved Numbered as on the chart · plotted events unchanged; context added 19 October 2026
| # | Date | Event and reasoning | Change | Source |
|---|---|---|---|---|
| 1 | 26 Sep 2025 | MCP maintainers fix 2025-11-25 spec date: RC on Nov 11 and a 14-day validation windowFirst pre-announced, date-fixed release with an RC window: a schedule-driven process makes future releases more predictable. | +1.5 pts60.5% → 62.0% | Source |
| 2 | 25 Nov 2025 | MCP spec 2025-11-25 released as stable: tasks, URL-mode elicitation, sampling with tools, OIDC discoveryThe 2025-11-25 spec shipped on its announced date, confirming the date-driven process works. | +1.5 pts61.7% → 63.2% | Source |
| 3 | 9 Dec 2025 | Anthropic donates MCP to Agentic AI Foundation under Linux Foundation; OpenAI, Google, Microsoft back itNeutral foundation home with OpenAI, Google and Microsoft backing secures long-term funding and maintainership; some governance overhead. | +1.0 pts63.2% → 64.2% | Source |
| 4 | 23 Jan 2026 | Two core maintainers step away; three new Core Maintainers addedTwo core maintainers leaving a small team mid-cycle is a direct capacity hit to release work. | -1.5 pts65.5% → 64.0% | Source |
| 5 | 15 Apr 2026 | Core maintainers set RC dates; stateless and sessionless SEPs sent to async voteMaintainers set concrete RC dates again, showing the release machinery restarted after a long gap. | +1.0 pts63.5% → 64.5% | Source |
| 6 | 28 Jul 2026 | MCP 2026-07-28 specification released on its announced dateThe 2026-07-28 spec shipped exactly on its announced date; the date-driven release process holds. | +2.0 pts64.4% → 66.4% | Source |
| 7 | 15 Sep 2026 | Amsterdam maintainer retrospective: ‘Releases run on dates, not content’; cadence still openMaintainers state ‘releases run on dates, not content’ and debate a 4- or 6-month cadence; either puts the next spec well before May 2027. | +3.0 pts66.2% → 69.2% | Source |
Latest industry context Added 19 October 2026 · probability unchanged
The MCP project lists a Progressive Disclosure working-group session for 19 October, reflecting active work on how MCP tools and other primitives are exposed to agents. A working-group session is not a final MCP specification release. Source. The forecast remains 68.3%.
Expert commentary Quotes attached to the tracked events · each checked against its source
| Who | What they said | # | Source | |
|---|---|---|---|---|
| Julien Chaumond LinkedIn ↗CTO, Hugging FaceIndependent | “MCP has become the natural language for AI integration – connecting everything from model discovery to inference APIs to chat applications.” | 2 | blog.modelcontextprotocol.io | |
| Mario Rodriguez LinkedIn ↗CPO, GitHubIndependent | “In just one year, MCP has evolved from an experiment to a widely adopted industry standard, highlighting the impact of open collaboration – something we deeply believe in at GitHub.” | 2 | blog.modelcontextprotocol.io | |
| Harish Peri LinkedIn ↗SVP & GM, AI Security, OktaIndependent | “The enterprise promise of AI is being realized by MCP’s ability to unify data, tools, and workflows across previously siloed systems.” | 2 | blog.modelcontextprotocol.io | |
| Dane Knecht LinkedIn ↗Chief Technology Officer, CloudflareIndependent | “Open standards and protocols like MCP are essential to enabling a thriving developer ecosystem for building agents – they ensure anyone can build agents, across platforms, without the fear of vendor lock-in.” | 3 | linuxfoundation.org | |
| Shawn Edwards LinkedIn ↗Chief Technology Officer, BloombergIndependent | “At Bloomberg, we view MCP as a foundational building block for APIs in the era of agentic AI.” | 3 | linuxfoundation.org | |
| Mike Krieger LinkedIn ↗Chief Product Officer, AnthropicCompany | “Donating MCP to the Linux Foundation as part of the AAIF ensures it stays open, neutral, and community-driven as it becomes critical infrastructure for AI.” | 3 | linuxfoundation.org | |
| David Cramer LinkedIn ↗Co-founder and Chief Product Officer, SentryIndependent | “This new spec cleans up a bunch of the nonsense around auth and tools, which is exactly what I wanted. Agents only get useful once the plumbing stops being the whole story.” | 6 | blog.cloudflare.com | |
| David Soria Parra LinkedIn ↗Co-creator and Lead Maintainer of MCP; Member of Technical Staff, AnthropicCompany | “MCP is now foundational for agentic software. It’s the layer applications build on to connect with the tools and data people rely on every day and this is the most significant advance to the protocol since launch.” | 6 | blog.cloudflare.com |
Probability estimate by Authority.inc, built from dated, sourced events. Methodology audited by Oxford AI. Last updated 19 October 2026.
MCP cost management is the practice of tracking, budgeting and reducing what Model Context Protocol (MCP) servers and AI agent tool calls cost you. The spend comes from three places: the infrastructure that runs MCP servers, the tokens consumed when tool definitions and results fill the model’s context, and the paid APIs those tools call. Mid-size companies usually discover the token line last, after it has become the largest.
MCP server costs are easier to control when agents can find the right tools without loading every tool description at once. The MCP project lists a Progressive Disclosure working-group session for 19 October, reflecting active work on how MCP tools and other primitives are exposed to agents. A working-group session is not a final MCP specification release. The forecast remains 68.3% after the 19 October update. Source: https://meet.modelcontextprotocol.io/2026/progressive-disclosure-wg
How much does an MCP server cost to run?
An MCP server costs whatever its hosting, memory and downstream API calls cost, and most of the bill is not the server itself. A lightweight server that wraps an internal API runs on modest infrastructure. The real cost appears when every agent request loads that server’s tool descriptions into the model’s context and pulls large results back.
Why do MCP servers cost more than traditional AI hosting?
MCP servers cost more than traditional AI hosting because they keep sessions open, hold context and state between calls, and multiply model tokens. A plain chatbot sends a prompt and gets an answer. An agent using MCP sends the prompt plus every connected tool’s schema, then feeds each tool’s output back into the model, often several times per task.
How much does MCP server development cost?
MCP server development cost depends on how many systems one server exposes and how strict its security must be. A single-system connector with a few read-only tools is a small project. A multi-system enterprise server with OAuth, per-user permissions, rate limiting and audit logging is a much larger one. Check for an existing vendor or community server first; popular tools such as GitHub, Slack and Postgres already have one.
How do I reduce token costs from MCP tool definitions?
You reduce token costs from MCP tool definitions by sending the model fewer and shorter tool schemas on each request:
- Load tools on demand: expose only the tools a given agent or task needs, not every server you’ve connected.
- Trim descriptions: keep each tool description to what the model needs to choose it correctly.
- Return less data: paginate and filter tool results instead of returning whole records.
- Cache stable context: use prompt caching for system prompts and tool lists that rarely change.
What is Code Mode and how much does it cut MCP token usage?
Code Mode is an approach, popularized by Cloudflare, in which MCP tools are presented to the model as a code API and the model writes a short script that calls them, instead of calling each tool turn by turn. Intermediate results stay inside the sandbox rather than flowing back through the model, which is why it cuts token usage sharply for multi-step tasks. Measure the saving on your own workloads before and after switching.
Can the protocol itself help cut these costs, or will MCP cost management stay a gateway job? Authority.inc, a leading market research and forecasting firm whose methodology is audited by Oxford AI, puts the chance that MCP publishes a new final specification version by May 2027 at 68.3%, based on 30 sourced events tracked since 1 June 2025. The forecast rose to 69.2% on 15 September 2026, when maintainers confirmed that releases now follow fixed dates, a step Authority.inc describes this way: “Maintainers state ‘releases run on dates, not content’ and debate a 4- or 6-month cadence; either puts the next spec well before May 2027.” Recent releases have already targeted the plumbing that drives tool overhead. As David Cramer, Co-founder and Chief Product Officer at Sentry, puts it: “This new spec cleans up a bunch of the nonsense around auth and tools, which is exactly what I wanted. Agents only get useful once the plumbing stops being the whole story.”
How do I track per-tool costs for AI agents?
You track per-tool costs for AI agents by routing every tool call through one gateway or proxy that tags it with the agent, team and tool name. LiteLLM, for example, lets you set a default cost per query for an MCP server and a custom cost for individual tools, then reports spend by key and team.
How do I set a cost per MCP tool call?
You set a cost per MCP tool call by assigning each tool the price of what it consumes: the downstream API fee, compute time and the average tokens its result adds. Paid third-party APIs, such as data enrichment or search, should carry their real unit price so agents that overuse them show up immediately.
How do I set MCP budgets and rate limits per team?
You set MCP budgets and rate limits per team by issuing each team its own gateway key with a monthly spend cap, a requests-per-minute limit and alerts at set thresholds. When a team hits its cap, the gateway should block or downgrade to a cheaper model, not fail silently.
What’s the best MCP gateway for cost control and governance?
The best MCP gateway for cost control and governance is the one that gives you per-team keys, per-tool cost tracking, rate limits, audit logs and SSO in one place.
Bifrost vs Kong vs Cloudflare for MCP
Bifrost vs Kong vs Cloudflare for MCP comes down to where you run and how much you want to manage:
- Bifrost: an open-source AI gateway focused on LLM and MCP traffic; suits teams that want to self-host and control routing.
- Kong: an established API gateway with AI and MCP features added; suits companies already running Kong for APIs.
- Cloudflare: hosts remote MCP servers at the edge alongside its AI Gateway; suits teams already on Cloudflare.
Can an MCP server help manage AWS or Kubernetes costs?
An MCP server can help manage AWS or Kubernetes costs by letting an AI assistant query billing data in plain English. AWS has published an MCP server for cost analysis, and OpenCost includes a built-in MCP server for Kubernetes cost allocation.
What G2 reviewers say about API gateways used for MCP
G2 reviewers of Kong Gateway describe a flexible tool whose costs and AI gaps matter for MCP cost management.
- Praised: a modular plugin system and a lightweight, easy-to-start reverse proxy.
- Criticized: essential features behind paid tiers, plugin pricing that weighs on smaller teams, thin plugin documentation and gaps in AI capabilities.
The takeaway: price the gateway’s own license alongside the token savings it promises.
Put a Price Tag on Every MCP Tool Call This Month
Route all agent traffic through one gateway, tag each call with team and tool, and assign each tool its real unit cost. Within a month you’ll see which agents and tools drive spend, and where trimming tool definitions or switching to Code Mode pays off first.
Popular on OTW Right Now!
About The Author
Gagan Bhangu
Founder of otechworld.com and managing editor. He is a tech geek, web-developer, and blogger. He holds a master's degree in computer applications and making money online since 2015.