Unified AI gateway for self-hosted infrastructure. Presents a single OpenAI-compatible API that routes to local inference or remote models based on model name, with conversation history, a database-driven bot framework, MCP tool integration, and a Ratatui TUI client.
  • Rust 80.2%
  • TypeScript 7.7%
  • Svelte 6.8%
  • Shell 3.6%
  • Python 1%
  • Other 0.7%
Find a file
Damien Coles 98ff3146af sync: mirror production at bdccb58
Session 16 work brought to public:

- Bucket A bug fixes: mid-stream cancel for single-turn agent runs,
  emitted_at on /v1/agent/runs replay, RunEmitter cancel flag wired
  to AgentRunManager.
- Bucket B-1 tool result cache (Valkey-backed, decorator over
  ToolDispatcher, per-server + per-tool TTL policy, degrades on
  Valkey outage). Spec at docs/specs/2026-04-24-tool-result-cache-design.md;
  plan at docs/plans/2026-04-24-tool-result-cache.md.
- TODO.md: tool cache + VRAM arbiter reclaim follow-ups.
- A/B harness fix: read .emitted_at instead of .payload.timestamp.
2026-04-25 06:43:59 -04:00
.idea sync: mirror production at 27f0dc2 + harden .gitignore 2026-04-23 19:33:48 -04:00
docs sync: mirror production at bdccb58 2026-04-25 06:43:59 -04:00
frontend sync: mirror production at 27f0dc2 + harden .gitignore 2026-04-23 19:33:48 -04:00
mcp-common sync: mirror production at 27f0dc2 + harden .gitignore 2026-04-23 19:33:48 -04:00
mcp-consul public-ready-init 2026-04-11 15:56:59 -04:00
mcp-local public-ready-init 2026-04-11 15:56:59 -04:00
mcp-loki public-ready-init 2026-04-11 15:56:59 -04:00
mcp-media sync: mirror production at 27f0dc2 + harden .gitignore 2026-04-23 19:33:48 -04:00
mcp-nomad public-ready-init 2026-04-11 15:56:59 -04:00
mcp-prometheus public-ready-init 2026-04-11 15:56:59 -04:00
mcp-web-search public-ready-init 2026-04-11 15:56:59 -04:00
oracle sync: mirror production at bdccb58 2026-04-25 06:43:59 -04:00
oracle-tui sync: mirror production at 27f0dc2 + harden .gitignore 2026-04-23 19:33:48 -04:00
tests sync: mirror production at bdccb58 2026-04-25 06:43:59 -04:00
.gitignore sync: mirror production at 27f0dc2 + harden .gitignore 2026-04-23 19:33:48 -04:00
BOT-FRAMEWORK.md public-ready-init 2026-04-11 15:56:59 -04:00
Cargo.lock sync: mirror production at bdccb58 2026-04-25 06:43:59 -04:00
Cargo.toml sync: mirror production at bdccb58 2026-04-25 06:43:59 -04:00
CHANGELOG.md public-ready-init 2026-04-11 15:56:59 -04:00
CLAUDE.md sync: mirror production at 27f0dc2 + harden .gitignore 2026-04-23 19:33:48 -04:00
deploy-config.sh sync: mirror production at 27f0dc2 + harden .gitignore 2026-04-23 19:33:48 -04:00
deploy.sh sync: mirror production at 27f0dc2 + harden .gitignore 2026-04-23 19:33:48 -04:00
oracle.toml sync: mirror production at bdccb58 2026-04-25 06:43:59 -04:00
README.md sync: mirror production at 27f0dc2 + harden .gitignore 2026-04-23 19:33:48 -04:00
TODO.md sync: mirror production at bdccb58 2026-04-25 06:43:59 -04:00

Oracle

AI gateway for the Arvandor platform. Unified OpenAI-compatible API routing to local inference (Ollama on RTX 4090) and remote models (Claude API).

Quick Start

# Build
cargo build --release -p oracle

# Run locally (needs DATABASE_URL and OATHKEEPER_SECRET)
cp .env.example .env  # edit with your values
cargo run -p oracle

Architecture

Oracle is a lightweight Rust/Axum proxy that:

  • Presents an OpenAI-compatible API (/v1/chat/completions, /v1/models)
  • Routes requests to Ollama or Claude based on model name
  • Authenticates via Ory Oathkeeper (browser) or API keys (CLI/agents)
  • Persists conversation history in PostgreSQL
  • Orchestrates tool-calling via MCP servers (Phase 3)

See CLAUDE.md for full documentation.