عن المشروع

فيليكس هو إطار عمل مستقل لاستضافة الوكلاء. يتم تأليف الوكلاء كمستندات YAML باستخدام apiVersion felix/v1؛ ويقوم المشروع بتركيبها إلى وكلاء خاضعين للحوكمة بمحاور مرنة (fibers)، ذاكرة، مهارات، تقييم، موافقات، وحارات عزل. exposes these agents through REST/SSE, an OpenAI-compatible /v1 API, A2A JSON-RPC, and MCP, and provides management APIs for audit, approvals, plans, jobs, manifests, eval, usage, and memory. Live runs can be forked, rewound, and steered, with session strategies for compacting, windowed, or semantic contexts. The stack includes an API service, a worker, and a scheduler. Postgres with pgvector is the system of record, with Valkey as cache and an object store that can be filesystem, S3, or GCS. Optional analytics warehouse backends include DuckDB, ClickHouse, and Doris. Deployment options include Docker Compose, Helm, AWS, and GCP, with a lean default image and optional extras for cloud SDKs, observability, and warehouse support. Governance features cover authentication, approvals, audit trails, usage meters, and retention controls. Observability uses Prometheus metrics and OpenTelemetry; model calls follow GenAI semantic conventions. Model providers include Anthropic, OpenAI, Ollama, Workers AI, and several OpenAI-compatible vendors, with logical model ids mapped by configuration. The model catalog holds context window, max output, price, parameters, thinking support, and modalities per family. Felix is designed for extension rather than dictating workflows. Plugins can register tools, HTTP routes, cron tasks, auth modes, rate-limit keys, body limits, self-authenticating mounts, startup hooks, audit/usage sinks, and agent-loop hooks. Open registries allow selecting patterns, model providers, object stores, secrets backends, warehouse backends, embedder backends, search backends, session strategies, and checkpointers through configuration. Agent Skills can be added as SKILL.md files without code. The repository includes bundled manifests such as quick, deep, router, oss-only, hybrid-router, support, cowork, governed, and contributor. A quick start copies .env.example, runs make install and make up, then uses the generated local API key against /chat or the OpenAI-compatible endpoint. The project states it runs on infrastructure the operator controls and is service- and cloud-agnostic, using Protocols rather than a single vendor SDK.