kastor

Terraform-style source-of-truth layer for AI agents using HCL specs

TLDR

Validate an agent specification

$ kastor validate examples/weather

Compile the spec to runnable LangGraph code

$ kastor build examples/weather

Plan changes for hosted agents (Terraform-style)

$ kastor plan examples/weather

Apply the plan to create or update agents

$ kastor apply examples/weather

Install via Homebrew or Go

$ brew install kastor

Kastor provides a declarative, versionable source of truth for AI agents. You define agents, tools, prompts, models and deployment targets in HCL files (`.agent`, `.tool`, `.prompt`, `kastor.hcl`). Kastor validates the spec, can generate runnable framework code (currently LangGraph), and supports plan/apply/destroy operations against targets with state, diffs and drift detection.It is not an agent runtime or executor; it manages the contract and reconciliation around agents.

PARAMETERS

Common options include paths to modules containing the declarative files. Many commands operate on a directory (the module root).

COMMANDS

validate [path]

Parse and validate references in a Kastor module.

build [path]

Compile the module into runnable framework code (e.g. LangGraph project).

plan [path]

Show a plan of changes against the target state.

apply [path]

Apply changes to create/update agents and state.

destroy [path]

Destroy agents managed by the module.

CAVEATS

- Early proof-of-concept; the CLI surface is stable but hosted provider support is limited.- Generated code is an output; the `.agent`/`.tool` files are the source of truth.- Requires appropriate credentials (e.g. OPENAIAPIKEY) when running generated agents.

RESOURCES

Source code · Homepage · Documentation