Product · SDKs

Govern the function, in Python or TypeScript.

When there's no MCP server and no model API to sit in front of, wrap the tool functions themselves. LangChain, OpenAI Agents, CrewAI or your own loop: the same policy, key, store and record.

Why it matters

The gap it closes

Some tools are just functions inside the agent's process. The SDKs put the same decision and the same signed record around them, and they're clear that this is the weakest insertion point.

What it does

Capabilities

Python

pip install decision-horizon. Decorate reads and consequential calls; a denial raises PolicyBlocked.

TypeScript

npm install @decision-horizon/sdk for Node 18+. Its records verify with the Python dh verify.

Same configuration

One dh.toml for policy, signing key and store, shared with the proxy and gateway.

Reads count

Wrapped reads feed the horizon, so a later write is judged against what the agent had seen.

Ostiary and upload built in

Both SDKs score each call with Ostiary and upload records to the hosted store when an API key is set.

Get started

In practice

From the documentation.

from decision_horizon.sdk import Session

with Session.from_config("dh.toml") as s:
    @s.read("read_file", ref_arg="path")
    def read_file(path): ...

    @s.govern("write_file", ref_args=["path"])
    def write_file(path, content): ...
Stated limits

What it doesn't do

  • The hook lives inside the agent process, so a compromised agent could call the function directly. Use the proxy wherever one can sit; the record states which insertion point it came from.

See it on one of your own workflows.

We'll put Decision Horizon in front of a real agent workflow of yours and replay the record with you.