Local AI runtime & agent orchestration
A Go-based local agent framework and web studio: a manager coordinates specialist bots through one resident Qwen3 model, with no cloud API required after setup.
Localbots makes a small local language model usable across several specialist roles on one laptop. A manager plans and reviews tasks, while a local web studio exposes progress, results, files, and follow-up requests. Go coordinates the runtime, llama.cpp serves Qwen3-1.7B, and SQLite stores goals, runs, traces, and checkpoints.
Shared one Qwen3-1.7B model served by llama.cpp, serialized model acquisition, and bounded queued work. Bots specialize through configuration and optional adapters rather than separate resident model copies.
Validated manager plans against a schema and real worker names, bounded retries and tool loops, and recorded task progress and checkpoints in SQLite.
Enforced granted filesystem roots in Go and isolated goal workspaces. File-tool grants are not a general sandbox for trusted executable plugins.
Built a separate QLoRA training/conversion path with pinned model identities, dataset checks, adapter manifests, and fixture comparisons. Reported synthetic results remain narrow experiments, not evidence of broad generalization.
A simplified view of the workflow.
Public source and repository evaluation records are available on GitHub. The runtime is Windows-first, and small-model outputs still need human review. Adapter results are limited to the documented experiments; no broad quality or performance gain is claimed here.
Tell me about the problem, the constraints, and what success would look like.
Discuss a similar project