daari
Local-first execution router — cache, tools, and local models before frontier APIs.
Status: v1.4.0 — docs site · tracker · release notes
Route Cursor, Claude Code, and any OpenAI-compatible client through local tiers instead of paying frontier for repeat work. Not a proxy — a cost optimizer you run. Apache 2.0.
Quick start
pip:
pip install daari
daari onboard --yes --serve
Docker (bundles Ollama):
docker compose up
First start pulls the L3 model (~2 GB), then daari listens on http://127.0.0.1:11435 (readiness: GET /ready). Prebuilt image: ghcr.io/naveenreddyalka/daari. Package: pypi.org/project/daari.
Docs — Developer documentation (install, concepts, guides, reference). Contributors: docs/DEVELOPING.md. Community: Discussions (templates).
One-click demo (install, serve, smoke curl, stats):
./scripts/demo.sh
Manual steps:
python3.12 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
ollama pull llama3.2:3b
daari serve
curl http://127.0.0.1:11435/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{"model":"llama3.2:3b","messages":[{"role":"user","content":"Say hi"}]}'
Run the same curl twice — the second response should show "tier": "L0" in daari_meta.
Feature snapshot (v1.2.0)
Routing & caching
- Local-first routing chain:
L0 exact cache → L1 semantic cache (with draft injection) → CCS/L2 rules → Lt CLI tools → L3–L5 local models (Ollama or MLX) → L6 frontier, with confidence-based escalation
- Opt-in compact-to-fit (
routing.compact_to_fit): trim droppable turns before L6; successful trims record tokens_before / tokens_after (stats: compact_to_fit_tokens_dropped)
- Cache trust you can measure: shadow-sampled false-hit rate, response-diversity monitoring, input normalization, per-category TTLs
- Prompt intelligence: category/complexity profiling, per-category policies, latency budgets, warm-model preference, learned routing from your own outcomes
Learning (on-device)
- Implicit outcome capture + explicit accept/reject feedback →
daari learn stats/recommend
- Auto-tuned per-category confidence thresholds, opt-in example capture →
daari learn finetune (MLX LoRA) → daari learn deploy
- Opt-in, review-first anonymized stats export (
daari learn export-stats) — metadata only, never prompts
Clients (one-click)
- Cursor (BYOK via tunnel + API-key auth), Claude Code (full tool passthrough), Claude Desktop (third-party gateway mode), ChatGPT Desktop / JetBrains AI Assistant (Ollama-compatible facade), VS Code, any OpenAI/Anthropic SDK
- Per-project profiles (
.daari.yaml): tier caps, no-frontier, latency budgets per repo
Observability & spend
- Per-request traces (
daari trace), usage ledger with estimated savings (daari report, Markdown export), web dashboard (daari web-ui serve)
- Monthly/daily frontier budgets with soft warnings, per-client cost attribution, optional pre-frontier PII scrub
Platform
- MCP server at
POST /mcp (initialize / tools/list / tools/call), Sourcegraph/GHE/GitLab providers, org shared cache + collective learning (tracer), gateway API-key auth, MLX backend for Apple Silicon
Docs
Principles
- Open source — Apache 2.0; run it, fork it, sell it
- Local-first — on-device by default
- Cost-minimize — cheapest capable path for every task
- AI optional — many tasks use IDE/CLI tools, not models
License
Daari is licensed under the Apache License 2.0. Decision: ADR-0016.
Repo
https://github.com/naveenreddyalka/daari
