Exact token counts

By default the size of skill texts is estimated as characters divided by four, and nothing leaves your machine. With --exact the name and description of every installed skill (the baseline listing) and the body of each skill are counted by the Anthropic count_tokens API instead:

export ANTHROPIC_API_KEY=...
agentlens --exact skills
agentlens --exact --exact-model claude-opus-5-5 unused

This is the only feature of AgentLens that sends text to a server, and it is off unless you ask for it. The application never uses it; it always estimates.

• Where it works: skills, context, unused, compare and report. Without ANTHROPIC_API_KEY the command stops before reading anything. --exact-model picks the tokenizer (default claude-sonnet-5-5); Claude models do not all tokenize alike.
• What is sent: the text of each installed skill (name plus description, and the body), one request per distinct text, to https://api.anthropic.com/v1/messages/count_tokens (or to ANTHROPIC_BASE_URL if set). Session contents are never sent. Skill files can contain private instructions; do not use --exact if they must not leave the machine. The counting endpoint is free of charge but rate limited.
• Cache: counts are stored by model and content hash in $XDG_CACHE_HOME/agentlens/count_tokens/<model>.json (~/.cache/...), so repeated runs send nothing for unchanged skills. Delete the file to start over.
• Marking: counted values are shown without ~. Values that are still computed stay estimates: carry, and the sizes of instruction files and tool definitions, which are known only by their file size.
• Failure: if the key is rejected, the network is down or the rate limit is hit, the run continues with the estimate, prints one warning on stderr and sends no further requests.
• The count includes the few tokens the API adds around a message, so it is slightly higher than the bare text.