v0.10.24 · Now in active development
Build agents that
think small, act smart
A unified ecosystem: Jeeves — the 96M param powerhouse model, Loomflow — the Python agent graph framework, and a purpose-built AI IDE to wire it all together.
0M
Parameters
0B tokens
Training data
0
Agent architectures
0+
Model providers
Jeeves Model
96 million parameters.
Infinite potential.
A compact instruction-tuned language model using Looped Transformer + Value Residual Learning. Trained on ~2B tokens — outperforms models trained on 20–150× more data.
Input tokens
→
Loop × N layers
→
Value Residual
→
Output logits
⚡
96M Parameters
Ultralight footprint. Runs on CPU, edge devices, and constrained environments without breaking a sweat.
🔁
Looped Transformer
Novel architecture using Value Residual Learning for deeper reasoning with a fraction of the parameters.
🛠
Tool Calling
Native function and tool-calling support baked in. Ready for complex agentic workflows out of the box.
💬
ChatML Format
Instruction-tuned with ChatML conversational format. Drop into any chat pipeline instantly.
🏆
20–150× Efficient
Outperforms models trained on 20–150× more data. The most compute-efficient model in its weight class.
🔓
Apache 2.0
Fully open source. Use commercially, fine-tune, redistribute — no strings attached.
Loomflow Framework
Agent graphs that
read like pseudocode
Two primitives. Agent hands the loop to the LLM. Workflow keeps you in charge of the graph. Drop either one inside the other.
support.py
⊞
Workflow Primitive
chain · route · parallel. Build the DAG yourself or use helpers. Cycles work; max_visits_per_node caps keep them honest.
⟳
12 Agent Architectures
ReAct, Plan-and-Execute, ReWOO, Reflexion, Self-Refine, ActorCritic, Tree of Thoughts, Router, Supervisor, Debate, Swarm, Blackboard.
◎
Multi-Tenant by Default
user_id is a typed primitive. Memory, budgets, permissions, audit logs all partition on it — no namespace leaks.
⬡
Composable Both Ways
Drop an Agent as a Workflow node. Or call wf.as_tool() — the entire workflow becomes callable from inside an agent.
◈
Model Agnostic
OpenAI, Anthropic, LiteLLM (~100 providers), Echo for tests. Swap with one kwarg. Native structured outputs supported.
⊡
Streaming First
agent.stream() and wf.stream() yield events with backpressure. Wire to SSE or WebSocket. Nothing buffers forever.
Coming Soon
The IDE built for
the agentic era
A purpose-built development environment where Jeeves and Loomflow are first-class citizens. Design agent graphs visually, run them live, iterate fast.
classify
route_billing
route_tech
agent.run()
wf.as_tool()
Jeeves Copilot
💡 Suggest adding error handler for billing agent timeout...
✓ Route graph validated. 2 agents, 3 nodes detected.
⚡ Jeeves local inference: 38ms avg latency
🧠
AI-Native Editor
Built from the ground up for LLM-assisted development. Not bolted on — integrated at the architectural core.
🔗
Loomflow Integration
Design and run agent graphs visually. Drag nodes, wire edges, watch execution propagate in real time.
🤖
Jeeves Copilot
Local completions powered by Jeeves-Small-95M. Fast, private, no cloud required.
📡
MCP Native
Connect any MCP server directly from the IDE. Tools, memory, retrieval — all wired through the graph.
🚧IDE in active development — join the waitlist to get early access