Every framework hands you a blank file. KuJI hands you the agent.

An open library of pre-configured agents for the jobs that look the same at every company — research, outreach, reconciliation, triage. Clone it, set one key, run it.

Opens at publication

Every agent framework on the market solves the interesting ten percent — state, tool calls, orchestration — and leaves the boring ninety to you: the prompt, the guardrails, the tool selection, the retry logic, the thing that stops it inventing an invoice number. A two-hour job becomes a two-week job, and most of them quietly never ship.

The jobs themselves are not unique. A lead-research agent at one company is ninety percent the same agent as at the next. KuJI is that ninety percent — packaged, evaluated and given away — so the two weeks go back into the ten percent that is actually yours.

One directory per agent

Prompt, tools, schema and evaluation cases ship together in a single folder. No framework lock-in and no configuration spread across six files.

One key to run

Set a provider key and run the command. No account, no dashboard, no telemetry, no hosted service standing between you and the model.

Readable before clever

Every preset is written to be opened and edited. If you cannot understand an agent in one sitting you should not be trusting it with production data.

MIT, permanently

Fork it, change it, ship it inside a product you charge for. No attribution theatre and no licence that changes once it gets popular.

Under the hood

Provider-agnostic, test-attached.

Presets declare their tools and their evaluation cases in the same place, so changing a prompt is a change that arrives with a test. Model choice is a line of configuration, not a rewrite.

# one directory per agent - prompt, tools, schema, evals$ kuji run research/lead-profile --input acme.com  ✓ 4 tools resolved   ✓ 12/12 evals pass

Built with

Python
TypeScript
MCP
Model-agnostic
MIT licence