Agentype

Agentype analyzes your local AI-agent usage and turns it into a private terminal overview.

It is for people who work across multiple agents and want one summary of their token footprint, usage rhythm, top projects, and preferred agents/models. With an opt-in LLM, it also uses project and skill context to infer a persona comment, keywords, and an archetype.

Supported Agents

Agentype currently reads local history from:

• Claude Code
• Codex
• OpenCode
• pi-agent
• Gemini CLI
• OpenClaw
• Nanobot
• Nanobot-compatible JSONL roots configured through AGENTYPE_NANOBOT_ROOTS

Availability depends on what each agent stores locally. Some agents provide richer token or tool metadata than others.

For Nanobot-compatible deployments outside the default ~/.nanobot/sessions path, set one or more roots with your platform path separator:

AGENTYPE_NANOBOT_ROOTS="/path/to/app:/path/to/another/app" agentype

Install

The PyPI distribution is agentype-cli because agentype is not available on PyPI. The installed command is still agentype.

The recommended CLI install is:

pipx install agentype-cli
agentype

If you do not use pipx, a regular pip install should also work:

pip install agentype-cli
agentype

For one-off runs without installing the command permanently:

uvx --from agentype-cli agentype

For source checkout development:

git clone https://github.com/cyzlmh/agentype.git
cd agentype
uv sync
uv run agentype

Usage

agentype
agentype -v
agentype --json-out
agentype --json-in output/agentype.json
agentype --png-out
agentype --output output --json-out --png-out
agentype --llm-base-url https://api.openai.com/v1 --llm-api-key sk-... --llm-model gpt-4o

• agentype prints the main terminal overview.
• agentype -v adds detailed statistics, discovered theme tables, and data confidence.
• agentype --json-out writes output/agentype.json.
• agentype --json-in PATH renders a previously written Agentype JSON file. This is meant for skill workflows where the triggering agent fills top-level persona fields before asking the CLI to print or render the final result.
• agentype --png-out writes output/agentype.png, a shareable poster-style summary.
• --output DIR changes where JSON and PNG artifacts are written.
• --llm-base-url URL --llm-api-key KEY --llm-model MODEL: opt into LLM persona discovery (BYOK, OpenAI-compatible API). All three are required; any can also be set via AGENTYPE_LLM_BASE_URL, AGENTYPE_LLM_API_KEY, AGENTYPE_LLM_MODEL.

What It Shows

• Persona poster: AGENTYPE and the LLM-inferred persona first, or a usage snapshot when no LLM is configured.
• Token usage: total, non-cache, cache, input, output, and reasoning tokens where available.
• Breakdowns: top projects, agents, and model IDs.
• Usage rhythm: monthly and weekly trend bars.
• Without LLM configuration, only the statistics overview is shown.
• When --llm-* options are provided (or the equivalent env vars), a persona comment, keywords, and an archetype inferred by the configured LLM from top project docs and skill metadata.

Agentype reports tokens, not cost. Pricing varies by provider, subscription, and plan, so cost estimates are intentionally excluded.

Privacy

Agentype is fully local by default. It reads known agent history locations on your machine and makes no network requests. If --llm-* options are provided, Agentype sends a compact usage summary (last-30-day token total, used agents, and used models), skill descriptions/categories/usage counts/canonical paths, and bounded top-project README.md, AGENTS.md, or CLAUDE.md text to the configured LLM provider. Raw prompts, transcripts, and detailed per-project/per-agent statistics are never uploaded in any mode.

Platform Support

Agentype targets macOS, Linux, and Windows. Release builds are checked with cross-platform CI.

Adapter coverage still depends on where each agent stores local history. Some agents may use different paths or formats across platforms.

If an agent stores its history somewhere different on your machine, Agentype may simply show less data instead of failing.

Agent Skill Usage

Agentype can also be packaged as an agent skill for marketplaces such as skills.sh and clawhub.ai. When triggered by an agent, it should run a workflow instead of stopping at the first statistics output.

Marketplace links:

• skills.sh: https://skills.sh/cyzlmh/agentype/agentype
• ClawHub: https://clawhub.ai/cyzlmh/agentype

Install from skills.sh:

npx skills add cyzlmh/agentype --skill agentype

Install from ClawHub:

npx clawhub@latest install agentype

Required skill loop:

agentype --json-out

The agent then reads the cached JSON at output/agentype.json, infers the persona/archetype from aggregate usage signals, and fills these top-level fields:

• archetype
• description
• keywords
• comment

After the JSON is filled, render the final result:

agentype --json-in output/agentype.json

For chat or IM gateway environments that can display images, also render the poster:

agentype --json-in output/agentype.json --png-out

Shell-oriented agents should return the rendered terminal summary. Chat or IM-oriented agents should send a compact summary and attach output/agentype.png when supported.

For nonstandard local agent history paths, configure available environment variables first. Nanobot-compatible roots use AGENTYPE_NANOBOT_ROOTS; unsupported layouts can be added in src/agentype/paths.py or a source adapter under src/agentype/sources/.