ClawAI

简介

   _____ _
  / ____| |
 | |    | | __ ___      __
 | |    | |/ _` \ \ /\ / /
 | |____| | (_| |\ V  V /
  \_____|_|\__,_| \_/\_/

Architecture Overview

Browser
  |
  v
Next.js Frontend (:3000)
  |
  v
Nginx Reverse Proxy (:4000)
  |
  +--> Auth (:4001) -------------> PostgreSQL claw_auth (:5441)
  +--> Chat (:4002) -------------> PostgreSQL claw_chat (:5442)
  +--> Connector (:4003) --------> PostgreSQL claw_connectors (:5443)
  +--> Routing (:4004) ----------> PostgreSQL claw_routing (:5444)
  +--> Memory (:4005) -----------> PostgreSQL claw_memory (:5445, pgvector)
  +--> File (:4006) -------------> PostgreSQL claw_files (:5446)
  +--> Audit (:4007) ------------> MongoDB claw_audit (:27018)
  +--> Ollama Service (:4008) ---> PostgreSQL claw_ollama (:5447)
  +--> Health (:4009) -----------> Aggregated service health
  +--> Client Logs (:4010) ------> MongoDB claw_client_logs (:27018)
  +--> Server Logs (:4011) ------> MongoDB claw_server_logs (:27018)
  +--> Image (:4012) ------------> PostgreSQL claw_images (:5448)
  +--> File Generation (:4013) --> PostgreSQL claw_file_generations (:5449)
  +--> Agent (:4015) ------------> PostgreSQL claw_agent (:5451)
  +--> Research (:4016) ---------> PostgreSQL claw_research (:5452)
  +--> Workspace (:4014) --------> PostgreSQL claw_workspace (:5450)
  +--> Llamacpp (:4017) ---------> PostgreSQL claw_llamacpp (:5440)
  +--> Payment (:4018) ----------> PostgreSQL claw_payments (:5453)
  +--> Threads (:4019) ----------> PostgreSQL claw_threads (:5458)
  +--> Thread Generation (:4020) -> PostgreSQL claw_thread_generation (:5457)
  +--> Plugins (:4021) ----------> PostgreSQL claw_plugin (:5459)
                                   + llamacpp-data volume (binary + GGUF weights)

Shared infrastructure:
  - RabbitMQ (:5672 / :15672)
  - Redis (:6380)
  - Ollama runtime (:11434)
  - ClamAV (:3310)
  - Prometheus, Grafana (behind the admin session), Vector log shipper
  - Optional research sidecars: Crawl4AI, FlareSolverr, Firecrawl

Flagship Features

1. Multimodal AI — voice and video notes, transcription, helper vision that describes images for text-only models, videos delivered to every model, and AUTO routing that picks a model able to handle the media. Send a file with no text and it is answered. 2. Files from chat — the AI writes PDF (typeset by Typst), Word, Excel, PowerPoint and Zip bundles, names them itself, within a daily allowance per plan. Download any answer in eight formats. 3. Smart attachments — ZIP, 7z, RAR, TAR and more expand as a tree, with a password retry in the chat; drop files anywhere; owner-only downloads of AI-made files that expire after an hour; ClamAV scanning on every upload. 4. Narrated AI research and crawling — an AI plans each turn's web work and narrates every step; pasted links are opened (up to 10 per message); site crawls up to 200 pages; fetches escalate from official APIs through TLS impersonation and a headless browser to optional Crawl4AI, FlareSolverr and Firecrawl sidecars and an archive snapshot — and robots.txt is respected. 5. Orchestration labs — Compare 2–5 models with a judge that ranks every answer, plus nine labs: Consensus, Escalation, Repair, Decompose, Best-of-N, Verify, Pipeline, Cost Ensemble and Role Pack — all with files and research. 6. Conversation power tools — branch from any message, edit and rerun, find in a conversation, export to Markdown, cross-thread context, prompt history. 7. Read aloud — metered, plan-gated text-to-speech that starts on the first segment and can be cancelled. 8. Images — generate with OpenAI, Gemini, xAI Grok or local Stable Diffusion and ComfyUI; shared conversation pages keep their images. 9. Pay-as-you-go credit — metered on 12 surfaces, prices shown in the visitor's currency while the charge settles separately. 10. 22 usable providers — OpenAI, Anthropic, Gemini, DeepSeek, xAI, Ollama, llama.cpp and 15 OpenAI-compatible presets (OpenRouter, Groq, Mistral, Kimi, GLM, Qwen, Cohere and more), the real model catalog, modality- and context-aware AUTO routing, and a shared provider circuit breaker. 11. Coding Agent for VS Code — a separate extension that signs in through the browser and obeys an organisation policy served by ClawAI. 12. Teams and administration — user management, roles and permissions, plan grants, per-user usage statistics. 13. Observability — an admin status page with uptime and incidents, Grafana and Prometheus, every container's log in one store, read-only ops tokens. 14. Local-first and privacy — Ollama, llama.cpp, local image generation, GPU overlay picked automatically, retention and memory controls. 15. Reliability — chat-service scales horizontally, dropped streams resume, Stop works across replicas, rolling deploys.

16. ClawAI Threads — turn a chat into a researched, cited public article: several models write, a Judge and a Critic review, you approve, then it is published with views, sharing and export (PDF, Markdown, JSON, TOON, HTML). Open to every plan; failures open an admin ticket. 17. Notifications and tours — a bell in the top bar and email tell you when a Thread is ready, published or failed (in your language); guided product tours for each page, offered once and switchable off everywhere. 18. Provider prices and model sync — models re-sync every six hours; OpenRouter prices are synced as provider-published rows (admin override > provider sync > seed); unpriced models are flagged, never recorded as zero. 19. Admin usage explorer — a tabbed Observability page (overview, consumption, AI and models, users, tracing, failures) with charts, searchable paged tables and a per-user consumption view; admin reads of user data are audited.

Also: memory and context packs, workspace connectors and automations, routing transparency on every answer, audit logging, and 13 interface languages with right-to-left support.

---