Keep agent behavior separate from production controls. You should be able to build a useful workflow locally, then add identity and a gateway Kill Switch without rewriting the agent.
If you’re shipping a product that calls models, read an agent quickstart with two questions in mind: what gets the workflow running, and where does that quickstart stop?
Stop Rebuilding the Agent Loop
Teams burn afternoons rebuilding the same agent execution loop before reaching the behavior that makes their product useful.
Kimss Forge is Kimss AI’s MIT open-source agent harness. It runs locally, is free, and requires no Kimss account. Start with a few lines and your model endpoint; keep your attention on the workflow rather than another harness implementation.
Install it:
pip install kimss-forge
For your first workflow, choose something small enough to inspect: one task, a narrow set of tools, and an outcome you can check.
Know What Belongs in Your Application
Your application should define what the agent is trying to do and which tools it can use. Production controls answer different questions:
- Which identity is making the model call?
- Where can access be cut off?
- Which requests actually passed through the gateway?
- When do organizational policies need to apply?
Forge gives you a local signal before you attach production controls: a soft Authority Boundary warning when a sensitive tool fires without a production gateway.
That warning never blocks local execution. It is a reminder to review the production boundary, not a substitute for gateway enforcement or the paid Authority Boundary feature.
Read the Benchmark Before Borrowing the Number
The Forge comparison in BENCHMARK.md reports about 80% less application code than the hard-way implementation.
Treat that as a result to inspect, not a promise about your codebase. Read it alongside the quickstart:
- What workflow was implemented?
- What counted as application code?
- Under which conditions was it tested?
- Does that resemble what you are building?
The useful takeaway is architectural: identify repeated execution plumbing you can remove, while keeping product-specific behavior explicit.
Attach Controls Without Rewriting the Agent
When you need the Kimss control plane, Forge’s attach path is:
gateway="kimss"
The same agent gains identity on model calls and a Kill Switch on the free Developer tier, when traffic uses the Kimss gateway.
The scope matters: the Kill Switch severs access at the gateway for routed traffic. It does not terminate local processes or control calls that bypass Kimss.
Authority Boundary, SSO, and SCIM are later, paid capabilities on the same path—not free Developer features, and not a reason to rebuild your agent.
Where the Gateway Fits
Kimss is a model-agnostic API gateway for production model calls, not a chat app or an editor assistant. You bring your model infrastructure; Kimss does not host models or resell compute.
For an existing OpenAI-compatible client, the routing change is a one-line base_url swap:
base_url="https://api.kimss.ai"
That is the routing setting, not a complete authenticated quickstart. Follow the gateway setup for credentials and your connected infrastructure.
Keep the distinction clear: Forge runs the agent; the gateway governs requests routed through it.
Start With One Inspectable Workflow
Build one small Forge workflow before adding more agent logic. Identify its sensitive tools, check the benchmark’s assumptions, and mark where the gateway would sit.
Next step: pip install kimss-forge.
If you ship with models, put a control plane in front of them—start with Forge.
Top comments (0)