LocalAgent

LocalAgent is a local coding backend for OpenCode and other OpenAI-compatible clients.

It combines:

• Go for the service layer
• Pebble-backed github.com/fgrzl/kv search and index storage
• Ollama for local models
• a workspace indexer that chunks files into searchable entities
• mux for the HTTP API layer

What it does now

• Scans a local workspace and indexes text files into Pebble through github.com/fgrzl/kv
• Exposes keyword search at /api/search
• Exposes OpenAI-compatible chat completions at /v1/chat/completions
• Injects relevant indexed workspace context into chat prompts automatically
• Advertises a curated chat model catalog through /v1/models and proxies Ollama embeddings

Environment

Defaults are chosen for local Docker use:

• LOCALAGENT_HTTP_ADDR=:8080
• LOCALAGENT_OLLAMA_URL=http://ollama:11434
• LOCALAGENT_WORKSPACE_DIR=/workspace
• LOCALAGENT_DATA_DIR=/data
• LOCALAGENT_CHAT_MODEL=qwen2.5-coder:7b-instruct
• LOCALAGENT_SUPPORTED_CHAT_MODELS=qwen2.5-coder:7b-instruct,qwen2.5-coder:14b-instruct
• LOCALAGENT_EMBEDDING_MODEL=nomic-embed-text
• LOCALAGENT_SEARCH_HIT_LIMIT=8
• LOCALAGENT_CONTEXT_CHUNK_LIMIT=4
• LOCALAGENT_REQUEST_TIMEOUT=90s

Run with Docker

docker compose up --build

Pull the chat models Rudy advertises before starting the app if they are not already present in your Ollama volume:

docker compose exec ollama ollama pull qwen2.5-coder:7b-instruct
docker compose exec ollama ollama pull qwen2.5-coder:14b-instruct
docker compose exec ollama ollama pull nomic-embed-text

The app container mounts:

• ./workspace for the code you want indexed
• ./data for Pebble data and the local index manifest

Model pulls

If you run Ollama outside Docker Compose, pull the models Rudy advertises before using OpenCode:

docker compose exec ollama ollama pull qwen2.5-coder:7b-instruct
docker compose exec ollama ollama pull qwen2.5-coder:14b-instruct
docker compose exec ollama ollama pull nomic-embed-text

OpenCode

OpenCode is the primary coding client for this repo. Point it at http://localhost:8080/v1 and use the same chat model as the service defaults.

See opencode/README.md for the exact setup.

Helpful endpoints

• GET /healthz
• GET /api/search?q=golang
• POST /api/index/rebuild
• POST /api/index/file
• POST /v1/chat/completions

Next steps

• Add file watching for incremental reindexing
• Improve snippets around matched text
• Add optional semantic/vector retrieval