DobiA plugin for Claude Code and Codex
Better results and lower token cost on the work you repeat with your agent.
Save a piece of work as a task. Dobi records every run, judges it against your checks and counts its tokens, so you can keep the version that does the job for less.
/plugin marketplace add https://github.com/run-llama/dobi.git /plugin install dobi@dobi
codex plugin marketplace add run-llama/dobi codex plugin add dobi@dobi
Needs Python 3.9 or newer on macOS or Linux. Nothing else to install. Install has the details.
How it works
Do the work once. Dobi keeps score from then on.
- 01
Do it once
Work with your agent as you always do. Then ask it to use Dobi to turn that into a task: what it's for, a few example inputs, and how to tell a good result from a bad one.
- 02
Record every run
Each time the agent runs the task, Dobi records the output, judges it against your checks and counts the tokens. Anything it couldn't see is marked unknown, never zero.
- 03
Keep the better version
Ask the agent to try a different prompt, model, tool or workflow. Dobi compares versions on the same inputs, so you can see which one does the job, and for how many tokens.
What runs where
- Your agent
- Does all the work, on your own subscription or API keys. Dobi's service makes no model calls, and it's never between your agent and a model.
- The plugin
- Runs on your machine and keeps its records in files only your user can read. It never sends the keys it keeps for you.
- The app
- Shows your tasks, runs and results. New tasks are Synced: their inputs, outputs and traces go to your account, redacted on your machine first. A Private task sends only IDs, hashes, verdicts, times and token counts. Private vs Synced says exactly what leaves your machine.