Glossary
The terms the CLI, the application and the MCP server use. The Glossary view of the application shows the same entries; both are made from one source.
Basics
Measured and estimated
A measured value was read from the client's own data, for example the token counts of a response. An estimated value was computed from file sizes, heuristics or the price table and is marked with a tilde (~) and the warning color. “n/a” means the client does not record the value at all.
Session
A session is one conversation with a client such as Claude Code, OpenCode, Copilot CLI or VS Code Copilot. Session time runs from the first to the last event, so idle time is included.
Request
Every time the client sends the conversation to the model, that is one request. Everything in the static context is sent again with each one, which is why small items add up.
Tokens
Tokens are the unit in which models read input and write output, and in which usage is billed. Where the client reports no count, AgentLens estimates tokens from the size of the content.
Source
A source is one supported client: Claude Code, OpenCode, Copilot CLI or VS Code Copilot. Use the source filter at the top to include or exclude them.
Subagent
Subagents work in their own context. Their requests are counted separately from the main agent's, which is why context size only covers main-agent requests.
Context
Static context
Tool definitions, instruction files (such as CLAUDE.md or AGENTS.md), agent and skill listings and MCP server instructions. The client sends them with every request, so they cost tokens whether they are used or not.
Context size
The number of tokens the model received for one request (uncached input plus cache reads and writes). It is measured and only counted for main-agent requests. It grows over a session until the client compacts it.
Overhead
A 500-token instruction file that was sent with 200 requests caused 100,000 tokens of overhead. Overhead shows which static content is expensive because it is large, because it is sent often, or both. It is an estimate. For skills it is baseline + load + carry.
Owner
The file, plugin, MCP server or built-in that provides the item. Group by project to see which project adds how much.
Kind
Tool listing and tool definition: the tools the model may call. MCP instructions: text an MCP server adds. Instruction file: CLAUDE.md, AGENTS.md and similar. Agent listing and skill listing: names and descriptions of installed agents and skills.
First request
Its size is measured, so it is an upper bound of all static context a session starts with.
Idle cost
What an installed skill, agent or MCP server costs while it does nothing. “Unknown” means the client reports no data for MCP servers.
Skills
Baseline
Clients list installed skills by name and description in each request so the model knows they exist. That is the baseline: the cost of having the skill installed, even if it is never used.
Load
When a skill is invoked, its full instructions enter the context once. Load is the size of that content.
File use
A skill is not only loaded by the client's skill tool or by typing its name: a model also reads its files and runs the scripts in its folder. AgentLens counts those tool calls per installed skill (only the skill name is kept, never the command or the path), and a skill with file uses is not listed as unused. Looking after a skill does not count: listing, printing, copying or committing its files, and editing them, and neither do the file accesses of a session that edits the skill. It is a signal, not proof. Loads are shown in the uses column; file uses are listed next to them.
Carry
Once loaded, the skill text stays in the conversation and is re-sent with every following request until a compaction removes it. Carry is that repeated cost and always an estimate.
Attributed
Only available from clients that report it (OpenTelemetry). Measured.
Follow-on
An assumption: everything that happens until something else takes over is attributed to the skill. Use it as a rough guide, not as a measurement.
Cost
Cost
Reported cost is shown as is. Otherwise the tokens are priced with the price table. Cache writes are priced as 5-minute writes. Add prices for unknown models to ~/.config/agentlens/prices.json.
Premium requests
Copilot counts requests against a monthly allowance instead of billing tokens. They are shown next to the cost, never added to it.
Unpriced
They are left out of totals and shares instead of being counted as free.
What if
Shows what the period would have cost on a different model. “Excluded” requests have no actual price and are left out of both sides.
Efficiency
Cache read
Repeated prefixes of a conversation are cached by the provider. Reading from the cache is much cheaper and faster than sending new input.
Cache write
The first time content is sent it is written to the cache. Later requests can read it.
Hit rate
Measured. “n/a” means the client reports no cache tokens.
Flow
Peak
Measured over the main agent's requests.
Near limit
Of the model's documented window, or, if the window does not fit the data, of the largest size at which that model was compacted.
Compaction
After a compaction, content that was carried along (such as loaded skills) is gone from the context.
Retry loop
Calls without a recorded outcome are “unknown” and never counted as errors.
Permission wait
It includes the time the user was away from the keyboard.