Devin MCP and token overhead
MCP (Model Context Protocol) connects Devin to external tools — Sentry, Datadog, Linear, Figma, databases, and hundreds more. On cloud Devin, MCP drives actions and context retrieval per session. On Devin Desktop, enabled MCP servers add tool schemas to agent context every turn even when no tool runs.
Hub: Tokenminning in Devin. Queries: devin mcp, disable MCP.
Cloud MCP marketplace
Cloud Devin uses the MCP Marketplace at Settings → Connections → MCP servers.
| Concern | Guidance |
|---|---|
| Enable only what the task needs | Each integration can trigger tool calls and context pulls |
| OAuth access mode | Organization = shared service account (recommended). Personal = per-user auth |
| Custom servers | Admins add STDIO, SSE, or HTTP transports — test with Test listing tools after save |
| Before the session | Connect Sentry/Datadog/DB MCPs in settings so the first message can reference them |
Cloud MCP cost shows up as actions and inference in session ACUs — not as a separate line item. A session that pulls large log payloads through Datadog MCP can spike ACU usage even with a short user message.
Trim strategy for cloud:
- List MCPs enabled org-wide — disable integrations no team uses this quarter
- Prefer org-level service accounts over personal OAuth for shared triage workflows
- Scope prompts to specific time ranges or resources so Devin does not dump entire log indexes
- Review Session Insights Issue Timeline for repeated failed MCP calls
Official examples: MCP use cases .
Desktop MCP (Cascade / Devin Local)
Each enabled MCP server in Devin Desktop adds schemas to every agent turn — even when no tool runs. Cascade enforces a 100-tool cap across all servers.
Disable MCP
When you are not using MCP this week:
- Turn off whole servers in Devin Desktop MCP settings
- Toggle off individual tools on each server (fewer schemas)
- Prefer one narrow server over five overlapping integrations
Devin Local vs Cascade
Devin Local prompts before MCP tool calls by default — fewer accidental invocations. Configure permissions once instead of approving every call.
CLI MCP
Devin CLI uses the same MCP configuration model — see CLI MCP overview . Imported configs from other editors may enable servers you do not need; audit after import.
Built-in vs MCP
| Need | Prefer |
|---|---|
| Local repo search | Built-in indexing / Fast Context |
| Remote GitHub issues | MCP GitHub server |
| Production logs | Cloud session + Datadog/Sentry MCP |
| Design specs | Figma MCP with scoped file links |
For local repos, built-in search often beats MCP GitHub/browser servers unless you need remote issue data.
Official Devin MCP servers
Cognition ships first-party MCP integrations:
- Devin MCP — query private and public repos from other agents
- DeepWiki MCP — architecture and docs context
Use these deliberately — they add capability but also tool surface area.
Troubleshooting
High input, low MCP use (Desktop) — schemas still attach. Disable servers.
Context overflow — MCP + always_on rules. Disable MCP first for fastest win.
Spike after enabling server — compare Plan Info before/after.
Cloud: high ACU, few messages — check Issue Timeline for MCP retry loops or oversized log fetches.