Kastor
Kastor is "Terraform for AI agents." Agents today are defined imperatively inside frameworks (LangGraph, CrewAI) or clicked together in platform UIs (OpenAI Assistants, Bedrock Agents) — there is no vendor-neutral, versionable, reviewable source of truth. Kastor provides one: a typed, declarative spec (.agent, .tool, .prompt files in HCL) and a Go toolchain with two paths — kastor build generates runnable projects for target frameworks, and kastor plan / kastor apply reconcile agents as long-lived resources on hosted platforms, with state, diffs, and drift detection.
The full design lives in SPEC.md.
Status
Kastor is an early proof of concept.
Working today:
parse .agent, .tool, .prompt, and kastor.hcl
validate references and prompt variables
build runnable LangGraph projects
examples: weather agent, content scheduler agent
Planned for v0:
kastor plan/apply
local state file and drift detection
Deploy to aws/azure platforms.
This is not another agent runtime/framework.
Install
Homebrew:
Install script (verifies the release checksum, installs to /usr/local/bin or ~/.local/bin, never sudo):
With Go 1.26+:
Or download an archive for your platform from the releases page, verify it against checksums.txt, and put the kastor binary on your PATH.
Quickstart: build the weather example
Prerequisites: Go 1.26+, Python 3.11+, an OpenAI API key, and a Tavily API key (the example's search tool runs against Tavily's hosted MCP server).
Compile the spec to a LangGraph project:
kastor build writes the generated project to examples/weather/gen/langgraph (the target's declared output). Generated output is not committed: it is reproducible from the spec, and codegen determinism is enforced by tests.
Set up the generated project:
The example's web_search tool is pinned to an MCP server and tool by its spec URI, mcp://search-server/tavily_search. How to reach that server is deployment configuration, not spec: create mcp_servers.json in the working directory (or point the KASTOR_MCP_CONFIG env var at a file elsewhere). For Tavily's hosted server:
The URL embeds your API key, which is why mcp_servers.json is gitignored — treat it as a secret, never commit it. Also note the spec URI's last path segment (tavily_search) must name a tool the server actually advertises, or calls fail with "does not expose tool".
Export the model credential (the example's model "fast" block uses provider openai):
Run the agent:
It prints the agent's declared output contract as JSON:
The generated README.md inside gen/langgraph owns the run-the-project side in full: every agent's inputs and outputs, tool bindings, and MCP configuration.
One v0 caveat (SPEC.md §3.2/§4): agent.weather's optional forecast_context input references agent.forecast's output. That reference is validated at compile time and orders the dependency graph, but generated code does not run the upstream agent for you — if you want the context, run forecast yourself and pass its summary via --inputs.
Development
SPEC.md is the source of truth for design decisions; CLAUDE.md documents the day-to-day conventions.