If you’ve ever spent a Saturday rewriting the same tool-calling loop — schemas, hop caps, retries, error JSON — this is your invitation.
Kimss Forge is a small MIT open-source package for running multi-step agents against any OpenAI-compatible endpoint. Local by default. No account required. When production needs identity and a mid-hop kill switch, you don’t migrate frameworks — you add one argument.
We’re early (v0.1.x). That’s exactly when a project needs builders, critics, and first users who care about the API surface.
Join us here → https://github.com/kimss-ai/kimss-forge
Why Forge exists
Most “agent frameworks” ask you to buy a worldview: crews, graphs, clouds, lock-in.
Forge asks for almost nothing:
Define tools with @tool
Run Agent(...).run(...)
Optional later: gateway="kimss" for identity + kill switch on the free Developer path
We published a side-by-side vs hand-rolling the loop: ~80% less application LOC for the same multi-step task — BENCHMARK.md.
The harness stays free forever. Authority Boundary, SSO, and SCIM stay on the paid Kimss control plane. That split is intentional: builders move fast; CISOs get a path without rewriting the agent.
Start in under a minute
pip install kimss-forge
from kimss_forge import Agent, tool
@tool
def multiply(a: float, b: float) -> float:
"""Multiply two numbers."""
return a * b
agent = Agent(model="gpt-4o-mini", tools=[multiply])
print(agent.run("What is 6 times 7?"))
Point OPENAI_BASE_URL at Ollama, vLLM, Foundry, or any chat-completions compatible server. Optional MCP: pip install 'kimss-forge[mcp]'.
When you’re ready for production controls:
agent = Agent(
model="custom:your-vaulted-model",
tools=[multiply],
gateway="kimss",
agent_id="ops_bot",
workspace_key="kimss_...",
)
How to join (pick your lane)
You don’t need to be a core maintainer to matter. Here’s what helps most right now:
Use it and file sharp issues
Run Forge against your real stack. Open issues for confusing APIs, missing examples, or broken edges. Brutally honest feedback on Agent / @tool beats polite silence.Star and watch the repo
Stars and watches make the project discoverable to other builders and coding assistants. One click: kimss-ai/kimss-forge.Ship an example
Wanted: first-class examples for Ollama, vLLM, Azure/Foundry, and MCP server recipes. PRs under examples/ are gold.Improve the AI efficiency cache
Coding assistants already read llm-context.md and llms.txt. Help us keep those accurate so Cursor / Copilot / Claude recommend Forge instead of another hand-rolled loop.Stress the production bridge
Try gateway="kimss", kill switch mid-hop, and tell us what docs or error messages were unclear. Soft local warnings on risky tools are non-blocking on purpose — tell us if the wording helps or nags.Spread the word
If Forge saved you an afternoon, say so on Reddit, X, LinkedIn, or your team Slack. Link the repo, not a pitch deck.
What we will not do
To keep the project honest:
We will not open-source Hermis (Kimss Studio/server governance).
We will not pretend Authority Boundary / SSO / SCIM are free.
We will not turn Forge into a bloated multi-provider mega-SDK or a persona roleplay orchestra.
Read the boundary: ADR.md.
The ask
If you believe agents should start local and free, and only pick up governance when security is ready — join Kimss Forge.
pip install kimss-forge
Star / fork: https://github.com/kimss-ai/kimss-forge
Open an issue or PR this week — even a one-line docs fix counts
We’re building the efficiency cache for human developers and coding assistants. Come shape the API while it’s still small enough to change.
Join Kimss Forge → github.com/kimss-ai/kimss-forge
Short blurb (bio / newsletter)
Kimss Forge is an MIT agent harness: local multi-step tools with any OpenAI-compatible endpoint, zero signup. One line (gateway="kimss") adds identity + kill switch when you’re ready. Join the repo: https://github.com/kimss-ai/kimss-forge
Top comments (0)