# Sources: Skills, MCP, and token use

The page separates measurements, scenario estimates, and illustrations. Skill, MCP, and Context Mode percentages must not be added because they can affect the same tokens.

## Local Context Mode aggregate

**Audit record:** `tests/fixtures/context-mode-aggregate.json`
**Access:** the aggregate is documented; raw session data is private.
**Supports:** Context Mode `1.0.162`; reviewed `2026-08-12`; `1,074` stats files; deduplication by `session_start`; `87` usable sessions; P25 `31%`, median `46%`, P75 `70%`, and `11/87` at `90%+` managed-output byte reduction. Token-equivalent uses `bytes / 4`.
**Does not prove:** exact provider tokens, provider billing, or whole-session savings. The displayed values were rounded independently: `23,783,102 + 3,369,455` is `16` above `27,152,541`.

## Public sources

### OpenAI Conversation state

**Link:** https://developers.openai.com/api/docs/guides/conversation-state
**Access:** public official documentation.
**Supports:** ways to preserve conversation state across turns.
**Does not prove:** our savings percentages.

### OpenAI Prompt caching

**Link:** https://developers.openai.com/api/docs/guides/prompt-caching
**Access:** public official documentation.
**Supports:** reuse of a prompt prefix.
**Does not prove:** logical context was removed.

### OpenAI Compaction

**Link:** https://developers.openai.com/api/docs/guides/compaction
**Access:** public official documentation.
**Supports:** compact conversation state.
**Does not prove:** preserved quality; that needs a separate check.

### MCP tools specification

**Link:** https://modelcontextprotocol.io/specification/2025-11-25/server/tools
**Access:** public specification.
**Supports:** tool contracts, schemas, pagination, and structured content.
**Does not prove:** that an MCP server saves tokens.

### Context Mode

**Link:** https://github.com/mksglu/context-mode
**Access:** public author repository.
**Supports:** the architecture for keeping heavy tool output outside the main context.
**Does not prove:** an independent provider-level A/B result for our system.

## Our Skill and MCP calculations

The common base is the measured `27,152,541` processed token-equivalent. Each range is calculated as `round(27,152,541 × percentage / 100)`.

| Scenario | Percentage scenario | Calculated token-equivalent |
|---|---:|---:|
| Unified workflow | `5–15%` | `1,357,627–4,072,881` |
| DevOps + SSH | `10–25%` | `2,715,254–6,788,135` |
| Systematic debugging | `10–30%` | `2,715,254–8,145,762` |
| Review + GitLab | `5–20%` | `1,357,627–5,430,508` |
| Production deploy | `5–15%` | `1,357,627–4,072,881` |

These values are arithmetic estimates on a shared managed-output base. They are not observed Skill savings, isolated MCP effects, whole-session savings, exact tokenizer output, or provider billing.

The GitLab MCP illustration repeats the combined Review + GitLab range. The SSH MCP estimate repeats the combined DevOps + SSH range. A provider-level A/B is still needed to isolate either MCP effect.