{"version":"2.1.319","anchor":"mcp-tool-images-are-counted-in-token-estimates","canonical_anchor":"mcp-tool-images-are-counted-in-token-estimates","heading":"Images from MCP tasks are counted as a fixed size when estimating conversation length","tier":"notice","area":"Context","scope":"individual","heads_up":false,"url":"https:\/\/changelogs.core-directive.com\/v\/2.1.319\/e\/mcp-tool-images-are-counted-in-token-estimates","release_url":"https:\/\/changelogs.core-directive.com\/v\/2.1.319","markdown":"### Images from MCP tasks are counted as a fixed size when estimating conversation length\n\nClaude Code now counts images arriving from MCP tasks as a fixed amount each when estimating tokens, not by the size of their text form\n\n**What**\n\nClaude Code keeps a running estimate of how many tokens, the units the model reads text in, a conversation uses. Images delivered by MCP tasks (work run by an MCP server, a plugin-like service that gives Claude extra tools) are now counted as a fixed amount per image. Before, they were counted by the size of their encoded text, which could be far larger.\n\n**Why**\n\nThe estimate decides when Claude Code compacts the conversation, meaning it summarizes older parts to free space. A closer estimate makes that happen at a better time when images come in from MCP tasks.\n\n- Area: Context\n- Tier: You'll notice\n- Useful: 2\/5\n- Signal: 1\/5\n- Scope: individual\n- Heads-up: no"}