Question 1

How many tokens does a single MCP server add to my agent's context window?

Accepted Answer

Each individual MCP tool costs 550–1,400 tokens for its name, description, JSON schema, field descriptions, enums, and system instructions. A single MCP server with 20 tools typically adds 8,000–15,000 tokens to every request. The tokens are consumed on every request, not just when the tools are used, because the model needs to know what's available at all times. At 10 servers with an average of 10,000 tokens each, you're looking at 100,000 tokens of fixed overhead. One developer reported a full MCP setup consuming 143,000 of 200,000 tokens — 72% of the context window — with MCP tools alone accounting for 82,000 tokens.

Question 2

What's the connection between MCP context bloat and the OpenClaw inbox wipe incident?

Accepted Answer

Context bloat makes compaction trigger sooner. Compaction is the process where OpenClaw compresses older conversation history when the context window fills. During the inbox wipe incident, compaction stripped a "confirm before acting" safety constraint, and the agent deleted 200+ emails. If MCP schemas are consuming 50% or more of the context window, the remaining space fills much faster, compaction fires more frequently, and any safety instructions in the conversation history face a higher chance of being compressed away. More MCP servers means less room, which means earlier compaction, which means safety constraints vanish sooner.

Question 3

Are MCP servers a security concern beyond token usage?

Accepted Answer

Yes. 30+ CVEs targeting MCP servers, clients, and infrastructure were filed in the first 60 days of 2026. One MCP package downloaded 437,000 times had a CVSS 9.6 remote code execution flaw. Each MCP server you add is a dependency with its own vulnerability surface — one that can expose your filesystem, network, system commands, and API credentials. The token cost and the security cost are two independent reasons to minimize your MCP server count. For the full picture, see our OpenClaw security guide.

Question 4

How do I know if my MCP setup is causing context bloat?

Accepted Answer

Count your active MCP servers and multiply by 10,000 tokens. If the result is more than 30% of your model's context window, you're in bloat territory. Symptoms include: frequent compaction during conversations (the agent "forgetting" earlier instructions), degraded reasoning quality on multi-step tasks, the agent selecting the wrong tool when multiple overlapping options exist, and slower response times as the model processes larger payloads. On a 128K model, 5 MCP servers already puts you at 39% overhead. On a 200K model, the threshold is around 6–7 servers.

Question 5

Can I load MCP servers dynamically instead of all at once?

Accepted Answer

Current MCP architecture loads all connected servers' schemas at connection time. You pay the token cost for every tool on every connected server, regardless of whether you use them in a given conversation. The practical workaround is to create separate agent configurations for different workflows — an email agent with Gmail MCP, a coding agent with GitHub MCP, a sales agent with CRM MCP — rather than one agent connected to everything. Teams are also building middleware that uses tool search to load definitions on demand and slices OpenAPI specs into smaller chunks. Cloudflare's Code Mode approach collapses 12 MCP round trips into 4 and delivers an entire API in roughly 1,000 tokens — a 90% reduction from traditional MCP schemas.

Question 6

Why do OpenClaw skills wrap MCP servers instead of using CLIs directly?

Accepted Answer

MCP provides a standardized protocol for tool discovery and invocation — the agent can inspect available tools, understand their parameters, and call them in a consistent way. CLIs are more token-efficient — Scalekit's 75 head-to-head comparisons showed MCP costs 4 to 32x more tokens than CLI for identical operations. But CLIs require the agent to know the exact command syntax and flags without schema guidance. Over 65% of active OpenClaw skills now wrap MCP servers because the protocol makes integration easier for skill authors. The tradeoff is token cost: easier integration for developers, more expensive context windows for users. The 2026 consensus is shifting toward a hybrid approach: skills for documentation, CLIs for local operations, APIs for external services, and MCP reserved for scenarios requiring dynamic tool discovery.

Question 7

What's a safe number of MCP servers to run simultaneously?

Accepted Answer

For most workflows on current models, 3–5 MCP servers is the practical ceiling before you start seeing compaction issues and tool-confusion degradation. At 3 servers (~30,000 tokens), you're using 15% of a 200K window or 23% of a 128K window — leaving plenty of room for conversation, memory, and reasoning. Above 7 servers, you're in compaction territory on 128K models. The right number depends on your model's context window and the token weight of the specific servers you're running, but fewer well-chosen servers consistently outperform more poorly-chosen ones.

Question 8

Are there alternatives to MCP that use fewer tokens?

Accepted Answer

Yes, and the benchmarks are clear. Scalekit's 75 head-to-head tests showed CLI uses 4–32x fewer tokens than MCP for identical tasks. Cloudflare's Code Mode collapses 12 MCP round trips into 4 and delivers an entire API in roughly 1,000 tokens. Sideko's benchmark across 12 Stripe tasks found Code Mode MCP used 58% fewer tokens than raw MCP. The emerging consensus for 2026: use skills for documentation, CLI tools for local operations, direct APIs for external services, and reserve MCP for scenarios that genuinely require dynamic tool discovery. The New Stack's "10 Strategies to Reduce MCP Token Bloat" is a practical starting point for teams looking to optimize.

Tag: MCP

MCP Context Bloat: How 10 MCP Servers Eat Half Your Agent’s Brain