tokensitterSee the context added by skills, plugins and MCP tool definitions. Then choose what deserves a place in every conversation.
Local context inventory · changes opt-inEvidence snapshot: 27 Oct 2026 · per-lever live deployment not verified
Availability differs between the API relay and the local plugin. Whether each lever is deployed on the live relay has not been independently checked.
The context study counts text from 29 public Claude Code plugins and 9 open-source MCP servers (149 tools) started locally; light, typical and heavy bundles of them. It measures carried context, not the usefulness of those tools.
The skills controls list plugins that add text to each conversation and can switch them to name-only or Off, with a backup. Check how each integration loads names.
The typical bundle is illustrative (Tokens always loaded by a typical setup (8 plugins, 3 MCPs)). Deferring schemas changes context cost; it does not establish that tools will still be selected correctly.
This experiment counted tokens and ran no response-quality tasks. MCPs & Skills: this is the cost of carrying the context, not the benefit of a skill or MCP you use. No tool-selection recall or downstream pass-rate claim follows from it.
Checks that would strengthen this evidence. They are not yet reported for this feature:
Test whether the relevant skill is found when its name alone is insufficient.
Verify required instructions are loaded before the skill is used.
Include tasks with several overlapping skills and distractor tools.
Measure classification and loading calls and cache effects, not just removed descriptions.
Opus 5.5 list prices: one 5-minute cache write, then 99 cache reads. Full MCP schemas versus deferred names. MCPs & Skills: the plugins are the public official set and the three setups are illustrative bundles, not a survey of real installs. The result is a floor estimate, not an invoice.
Results are copied from our lab reports (snapshot 27 Oct 2026) and were not re-run for this website.
| Experiment | Workload and date | Result | Limits | Report |
|---|---|---|---|---|
| MCP and plugin context count | 29 public Claude Code plugins and 9 open-source MCP servers (149 tools) started locally; light, typical and heavy bundles of them 26 Oct 2026 | 19,033 Tokens always loaded by a typical setup (8 plugins, 3 MCPs) $0.47 Typical setup (8 plugins, 3 MCPs): cost per 100-turn session of the context it always loads (MCP tool names only: $0.12) $0.12 Typical setup (8 plugins, 3 MCPs) with MCP tool names only, cost per 100-turn session 33,442 Tokens the Notion MCP server (24 tools) adds to every conversation $1.98 Cost per 100-turn session of a heavy setup (all 29 plugins, 9 MCPs) $0.33 The same heavy setup with MCP tool names only | Token-count study only; no response or task-quality test. MCPs & Skills: the plugins are the public official set and the three setups are illustrative bundles, not a survey of real installs. |
|
Lab results at Anthropic API list prices on fixed tasks show what a lever did on those tasks, not what it will do on your bill. For plugin levers on Pro or Max, list-price values describe usage economics, not the subscription invoice. Relay levers apply to API keys only.
Read the measurement method and evidence limitsThis study does not test that. It counts tokens. Selection recall, instruction loading and task success need separate evaluations.
No. They are list-price cost illustrations for one fixed session under the stated context and cache assumptions.
Add Tokensitter to Claude Code and see where your tokens go.