Local Model Runtime
Model routing supports local/open model paths such as Ollama, llama.cpp, MLX, and Hugging Face GGUF download planning with license and local-fit checks.
Local Claude-style AI development partner
Aster turns local open models into a coding partner with memory, tests, review, and learning data.
Aster is an Electron IDE for private, local-first agentic coding. It opens a real project, gathers context, proposes patches, runs checks, reviews evidence, and turns useful traces into auditable SFT/DPO dataset candidates.
262
core tests passed in latest full harness run
192/6
harness pass/fail history in uploaded summaries
21
redacted trace records in the local dataset export
2
fine-tuning rows from one grounded SFT and one DPO pair
What Aster Is
Aster is not a clone of Claude or Anthropic private systems. It is a local-first product that recreates the developer workflow people want: context, tool use, validation, memory, and data ownership.
Multi-Agent Loop
Planner, Researcher, Coder, Tester, Reviewer, and Judge roles create a feedback loop around patch proposals, test evidence, review blockers, and retry traces.
Memory And Learning
Context packs, LM Wiki, graph memory, review queues, and exported traces create a path from daily work to private SFT/DPO/LoRA data.
Evidence Already Shipped
The private dataset now contains a trace snapshot, review decisions, sanitized harness summaries, and the first fine-tuning candidates from a real local LLM Agent Loop.
Repo
Private GitHub repo with Electron, React, FastAPI, agent loop, model router, and tests.
Dataset
Private Hugging Face dataset with trace snapshot, review decisions, manifest, and card.
Harness
198 sanitized run summaries without local paths, commands, stdout, or stderr.
Gate
Dataset manifest marks Training ready as true after grounded SFT and DPO rows appeared.
Next
Scale from one verified local loop to repeated real-project SFT/DPO candidates.
Next Milestone
Step 1
Loaded a local Ollama model and verified Aster can route Agent Loop roles to it.
Step 2
Ran a real Agent Loop task on a small repo: patch, test, review, and judge passed.
Step 3
Promoted one grounded SFT row and one DPO pair into the local dataset export.
Step 4
Next: repeat on real codebases and push the stronger dataset snapshot after review.
Honest Risks
Model Quality
Local output quality depends on model choice, quantization, runtime, memory, and eval gates.
UX Polish
The technical loop exists, but first-run onboarding and model download flow still need polish.
Training Data
The first candidates exist. The next risk is scale: quality review must stay strict as rows grow.
Private local AI coding should own the workflow layer.
As open models improve, Aster's durable advantage is the local IDE loop around code, memory, validation, review, and user-owned learning data.