← All articles

How Many MCP Servers Should You Run?

June 17, 2026·16 min read·MCPForge

Key Takeaways


Start with one MCP server only if the tools share the same risk level, owner, permission model, and operational lifecycle. Split MCP servers when tools belong to different business domains, teams, credentials, approval requirements, or security levels. Do not organize MCP servers only by technical convenience. Organize them by trust boundaries.


A large "everything server" becomes difficult to govern, audit, approve, and safely expose to AI agents. Too many tiny MCP servers create discovery, deployment, monitoring, and maintenance overhead. Production MCP architecture should balance developer experience with security team requirements.


Tool permissions, approval workflows, audit logs, credentials vaulting, side effect detection, and security reports become more important as the number of MCP servers grows.


The Simple Answer


For most teams, the right starting point is:


Team stageRecommended MCP server count
Prototype1 MCP server
Internal pilot1–3 MCP servers
Production single team2–5 MCP servers
Production multi-team5–20 MCP servers
Enterprise platform20+ MCP servers, governed centrally

This is not a rule. It is a starting point. A small startup might run one MCP server for months. A large enterprise may need separate servers for engineering, finance, support, sales operations, security operations, and internal data access from day one.


The number matters less than the boundary. One well-governed MCP server is better than ten unmanaged ones.

Want to analyze your API security?

Import your OpenAPI spec and generate a Security Report automatically.

Generate Security Report →

Why This Question Exists


Traditional APIs are usually consumed by deterministic software. MCP changes the access pattern. With MCP, tools are exposed to AI agents that can reason, select tools, chain actions, and combine data from multiple systems.


An MCP server is not just a technical wrapper around an API. It is a capability surface for an AI system.


ToolTraditional API concernMCP concern
get_customerIs the endpoint authenticated?Should this agent see customer data at all?
create_refundIs the user authorized?Should an AI agent trigger refunds without approval?
delete_projectIs the request valid?Should this tool exist in the agent context?
send_emailDoes the email service work?Could the agent send the wrong message to a customer?
deploy_serviceDoes CI/CD accept the command?Could the agent deploy unsafe code?

This is why MCP server count becomes a governance decision, not just an infrastructure decision. See [what is an MCP server](/blog/what-is-an-mcp-server) for a deeper explanation of the MCP model.


The Wrong Way to Decide MCP Server Count


The most common mistake is organizing MCP servers around convenience: "Put everything in one server because it is easier," "Create one server per API," "Create one server per team," "Create one server per agent."


The better question is: **Which tools should share the same trust boundary?**


A trust boundary includes who owns the tools, who approves access, what credentials are used, what data can be accessed, what actions can be performed, what audit logs are required, and what compliance requirements apply.


What an MCP Server Boundary Really Means


A good MCP server boundary makes it easier to answer: What can this server do? Which systems does it touch? Which agents are allowed to use it? Which tools require approval? Where are its audit logs? Who owns incidents caused by this server?


If nobody can answer those questions quickly, the server is too broad or too poorly governed.


Pattern 1: One MCP Server for Everything


This is the fastest way to start. It can work for prototypes, demos, and very small internal experiments. It becomes risky in production.


Advantages: simple deployment, easy discovery, fast iteration, lower operational overhead.


Disadvantages: weak separation (read-only and dangerous tools live together), harder security review, credential risk, permission complexity, audit noise, and large blast radius.


**One MCP server is acceptable when:** the project is experimental, tools are read-only, one team owns all tools, one credential scope is enough, and there are no destructive operations.


Pattern 2: One MCP Server per Business Domain


This is usually the best production starting point. Instead of one server for everything, you split by domain:


engineering-mcp-server → search_repos, create_github_issue, read_ci_status

support-mcp-server → search_customers, get_ticket_history, draft_reply

billing-mcp-server → get_invoice, get_subscription, request_refund_approval


Domain-based splitting usually works because ownership, permissions, credentials, and risk often follow business domains. A security team can review the billing MCP server separately from the engineering server. IAM can assign different access policies.


**Use this model when:** multiple teams own different tools, tools touch different business systems, access should map to business roles, some domains contain sensitive data, and you expect tool count to grow.


For most production teams, this is the right default. See [MCP security best practices](/blog/mcp-security-best-practices) for governance patterns.


Pattern 3: One MCP Server per Risk Level


Another strong model is splitting by risk:


readonly-mcp-server → search_docs, get_customer_summary, list_invoices

controlled-actions-mcp-server → create_ticket, draft_email, schedule_report

critical-actions-mcp-server → issue_refund, delete_account, deploy_service


Risk levelDescriptionExample tools
Read-onlyRetrieves data without changing stateSearch docs, list invoices
Low side effectWrites harmless or reversible dataAdd note, create draft
Medium side effectChanges workflow stateCreate ticket, update status
High side effectFinancial, operational, or customer impactRefund, cancel plan
Critical side effectIrreversible or privileged operationDelete account, deploy production

**Use this model when:** security review is strict, you expose tools to autonomous agents, you have destructive or financial operations, and approval workflows are required.


Pattern 4: One MCP Server per API or Vendor


Many teams naturally create one MCP server per API. This feels clean because it maps directly to existing systems. It is sometimes useful, but a vendor boundary is not always a trust boundary.


For example, Stripe contains both low-risk read tools and high-risk financial tools. In production, you may need `stripe-readonly-mcp-server`, `stripe-billing-actions-mcp-server`, and `stripe-critical-actions-mcp-server` — or domain-based equivalents.


Pattern 5: One MCP Server per Agent


Some teams create one MCP server per AI agent. This can work when agents are stable products with clear responsibilities. However, it can become hard to maintain if agents change often, and tool duplication across servers becomes a governance problem.



For most teams, the best production architecture is a hybrid:


  • Split MCP servers by business domain
  • Separate high-risk actions from read-only tools
  • Use tool-level permissions inside each server
  • Require approval workflows for critical side effects
  • Keep audit logs across all servers
  • Store credentials in a vault
  • Generate security reports before production exposure

A practical production layout:


docs-readonly-mcp-server

engineering-readonly-mcp-server

engineering-actions-mcp-server

support-readonly-mcp-server

support-actions-mcp-server

billing-readonly-mcp-server

billing-approval-actions-mcp-server

admin-critical-mcp-server


How Tool Count Affects MCP Server Design


Tool count matters because AI agents perform better when tools are understandable and relevant.


Tools per MCP serverRecommendation
1–5Usually too small unless tools are high-risk
5–20Healthy range for focused servers
20–50Acceptable if tools share one domain
50–100Requires strong naming, permissions, documentation
100+Usually a sign you should split the server

Tool sprawl is a real production problem. When too many tools are exposed, agents may select the wrong tool, misunderstand overlapping tools, or combine tools in unexpected ways.


How Risk Should Influence MCP Server Count


The more dangerous the tools, the more separation you need.


A useful rule: Read access can be grouped more broadly. Write access should be split more carefully. Critical side effects should be isolated.


Side Effect Detection and Why It Matters


A side effect is any tool action that changes state outside the agent — creating a ticket, sending an email, updating a database row, charging a card, triggering a deployment, changing permissions.


Before deciding how many MCP servers to run, classify every tool by side effect:


CategoryDescriptionExampleServer strategy
No side effectReads data onlyget_invoiceCan be grouped
Internal side effectChanges internal metadataadd_noteUse permissions
External side effectAffects customers or vendorssend_emailSeparate or approval-gate
Financial side effectChanges money, billingcreate_refundIsolate
Privileged side effectChanges access or infrastructurerotate_keyIsolate heavily

See [how to secure an MCP server](/blog/how-to-secure-an-mcp-server) for a complete governance framework around side effect detection.


Permissions: Server-Level vs Tool-Level


If you only have server-level permissions, you need more servers. If you have strong tool-level permissions, you can safely group more tools.


Tool-level permissions answer whether a specific agent or user can access a specific tool:


Agentget_customerlist_invoicescreate_refunddelete_customer
Support assistantYesYesNoNo
Billing assistantYesYesApproval requiredNo
Admin assistantYesYesApproval requiredApproval required

Tool-level permissions reduce the need to split every risk difference into a new server — but they do not remove the need for server boundaries.


Approval Workflows Change the Architecture


Approval workflows allow agents to request actions without executing them immediately. This makes some tools safer to expose. But a better model is layered governance: read-only servers (no approval needed), low-risk action servers (limited approval), high-risk action servers (human approval required), critical admin servers (restricted access plus approval).


Audit Logs Become Harder With Too Many or Too Few Servers


If you run one giant server, audit logs become noisy — support, engineering, billing, and admin events are all mixed. If you run too many tiny servers, audit logs become fragmented.


Each audit log entry should include: MCP server name, tool name, agent identity, user identity, input parameters, output summary, risk classification, approval status, credential used, timestamp, result, and correlation ID.


Credentials and Vault Strategy


Credential boundaries are one of the strongest signals that you need separate MCP servers. If two tools require completely different credentials, scopes, or secret rotation policies, they may belong in separate servers.


If an MCP server needs credentials for ten unrelated systems, it is probably too broad.


Decision Framework: Should You Split an MCP Server?


Split the MCP server if the answer is "yes" to any of these:


QuestionSplit?
Do tools have different business owners?Usually yes
Do tools use different credentials?Often yes
Do tools have different risk levels?Often yes
Do some tools require approval?Usually yes
Do tools access regulated data?Usually yes
Do tools perform destructive actions?Yes
Do tools affect billing or money movement?Yes
Do tools affect production infrastructure?Yes
Is the tool list becoming confusing?Maybe
Are audit logs too noisy?Maybe
Has the server no clear owner?Yes

Do not split only because the file is getting large. Split because the trust boundary changed.


Naming Conventions for MCP Servers


Use names that communicate domain and risk:


billing-readonly-mcp-server

billing-actions-mcp-server

support-readonly-mcp-server

admin-critical-mcp-server

customer-data-restricted-mcp-server


Avoid: `tools-server`, `internal-mcp`, `agent-server`, `api-mcp`.


A good MCP server name answers: what domain, is it read-only or action-oriented, is it restricted, who probably owns it.


The Danger of Raw Database MCP Servers


One of the most dangerous patterns is exposing raw database access through MCP with tools like `run_sql_query`, `update_table`, `delete_rows`. This is hard to classify for side effects, hard to audit for intent, and easy for agents to misuse.


A safer model exposes task-specific tools: `get_customer_by_email`, `list_recent_failed_payments`, `get_subscription_status`.


Environment Separation


Do not mix development, staging, and production in one MCP server unless strong environment boundaries are enforced. Production write tools require stricter permissions, production critical tools require approval, and production credentials must be scoped.


Common Mistakes


**Mistake 1: One server because it is faster** — fine for prototypes, dangerous in production without review.


**Mistake 2: One server per OpenAPI spec automatically** — OpenAPI specs are technical artifacts; MCP servers are trust boundaries.


**Mistake 3: Mixing read-only and destructive tools** — forces security teams to treat the whole server as dangerous.


**Mistake 4: No tool-level permissions** — without them, server count must carry too much responsibility.


**Mistake 5: No approval workflow** — some tools should not execute immediately just because an agent selected them.


**Mistake 6: No audit logs** — if you cannot reconstruct what the agent did, you are not ready for production.


**Mistake 7: No owner** — every MCP server should have an owner responsible for tools, credentials, documentation, incidents, and reviews.


Final Recommendation


Most teams should not ask "How many MCP servers should we run?" They should ask: **What are our MCP trust boundaries?**


Start with one server for prototypes. Move to domain-based servers for production. Separate read-only tools from high-risk actions. Isolate critical operations. Use tool-level permissions, approval workflows, audit logs, credentials vaulting, side effect detection, and security reports as your MCP footprint grows.


A good MCP architecture is not the one with the fewest servers. It is the one where every server has a clear purpose, clear owner, clear risk level, clear permissions, clear audit trail, and clear reason to exist.


[Import your OpenAPI spec](/login?tab=register) and generate a Security Report to understand your current MCP risk surface before deciding how to split servers.

Frequently Asked Questions

How many MCP servers should a small team run?

A small team should usually start with one MCP server for prototypes and one to three servers for production. The first split should usually separate read-only tools from tools that change data or trigger external actions.

Should every API become its own MCP server?

No. An API boundary is not always a trust boundary. Some APIs contain both safe read tools and dangerous write tools. MCP servers should be organized around ownership, risk, credentials, permissions, and governance.

Is one large MCP server bad?

Not always. One large MCP server can work for read-only internal tools or early prototypes. It becomes risky when it includes unrelated domains, high-risk actions, multiple credentials, regulated data, or unclear ownership.

When should we split an MCP server?

Split an MCP server when tools have different owners, credentials, risk levels, approval requirements, compliance needs, or audit requirements. Also split when the tool list becomes too large for agents and developers to understand.

Should read-only and write tools be in separate MCP servers?

Usually yes, especially in production. Read-only tools can often be exposed more broadly, while write tools need stricter permissions, approvals, and audit logs.

How many tools should one MCP server expose?

A focused MCP server often has 5–20 tools. Servers with 20–50 tools can work if they share one domain. Servers with more than 50 tools need strong documentation, permissions, and review. More than 100 tools is often a sign that the server should be split.

Should each AI agent have its own MCP server?

Sometimes, but not by default. Agent-specific servers work when agents are stable and narrowly scoped. For most teams, domain-based and risk-based server boundaries are easier to govern.

How should enterprises manage many MCP servers?

Enterprises should use a central MCP registry, tool inventory, ownership metadata, risk classification, tool-level permissions, approval workflows, audit logs, credentials vaulting, and security reports before approving production exposure.

What is the biggest risk of running too many MCP servers?

Too many MCP servers can create discovery problems, duplicated tools, fragmented monitoring, inconsistent security policies, and operational overhead. Splitting should follow real trust boundaries, not arbitrary technical preferences.

What is the biggest risk of running too few MCP servers?

Too few MCP servers can create large blast radius, unclear ownership, mixed credentials, noisy audit logs, and difficult security reviews. A single server containing billing, admin, support, and engineering tools is usually too broad for production.

Should dangerous tools be exposed through MCP at all?

Some dangerous tools can be exposed if they are isolated, permissioned, approval-gated, audited, and carefully described. But not every API endpoint should become an MCP tool. Some operations should remain manual or require separate internal systems.

What is the best default MCP architecture?

The best default production architecture is domain-based splitting with risk-based separation. Start with business domains, separate read-only from action tools, isolate critical side effects, and apply tool-level permissions and approval workflows.

Check your MCP security posture

Generate a Security Score, detect risky tools, and review permissions before exposing APIs to AI agents.

Try MCPForge Free →