Your coding agents already write everything down. Claude Code, Codex, Cursor, Antigravity, OpenCode and sixteen others leave JSONL transcripts in your dotfiles and never look at them again.
AgentWorth reads those files, indexes them into a local SQLite database, and grades what each session claimed against what it left behind. Nothing is uploaded. There is no account and no server.
Install
Pick one. The first needs nothing installed.
npx -y agentworth scan
curl -fsSL https://agentworth.dev/install.sh | sh
cargo install agentworth-cli
archie is a shorter alias for agentworth. Every command below accepts either
name.
The first scan
archie scan
scan walks the adapter directories, fingerprints each source file, and indexes
what changed. Run it again tomorrow and it only reads the files that moved.
--force re-reads everything, which you want after an upgrade that adds a new
field.
If nothing shows up, archie doctor prints what it looked for and what it
found.
The first three commands
Where the tokens went.
archie stats
One screen: sessions indexed, adapters detected, tokens in and out, cache reads, estimated spend, and the highest outcome rung reached.
Which sessions exist.
archie session list --limit 20
archie session list --adapter claude_code --model opus
The newest sessions, with their repo, model, event count and outcome. Filter by adapter or by a substring of the model name.
The explorer.
archie serve --open
A local server on port 3000, and a browser tab pointed at it. Prompts, thinking blocks, tool calls, file diffs, compaction rounds — the whole session, step by step.
Where to go next
- The outcome ladder — what "verified" means here.
- The handoff — what a session promised, decided and dropped.
- MCP setup — let the next session ask, instead of you.
What to run
npx -y agentworth scan
archie stats
archie session list --limit 20
archie serve --open
archie doctor