freshcrate
Skin:/
Home > MCP Servers > seraph

seraph

An AI guardian that remembers, watches, and acts.

Why this rank:Release freshnessHealthy release cadenceStrong adoption

Description

An AI guardian that remembers, watches, and acts.

README

Seraph

An AI guardian that remembers, watches, and acts

Tests Python 3.12 React 19 MIT License

Seraph is a proactive AI system for people who want an agent that can keep state, watch what is happening on the desktop, use tools, run workflows, and surface useful actions instead of waiting for one-off prompts.

seraph-demo.mp4

Watch the 20-second workspace demo


What Seraph Is

Seraph is a workspace-first AI agent with:

  • persistent identity and long-term memory
  • proactive scheduling, briefings, reviews, and intervention policy
  • tool use, reusable workflows, and plug-and-play MCP server integration
  • a browser cockpit for live operation and inspection
  • an optional macOS daemon for window tracking and OCR-backed screen awareness

This repository is for builders and power users who want to run, inspect, and extend a serious guardian prototype rather than a chat-only assistant shell.

Current State

Shipped today on develop:

  • browser workspace UI, backend APIs, observer daemon, memory, goals, and proactive scheduler foundations
  • 17 built-in tool capabilities plus workflow, starter-pack, skill, and MCP integration surfaces
  • runtime routing, fallback, approval, audit, and policy foundations
  • a dense operator cockpit with activity history, approvals, interventions, and workflow inspection

Still in progress:

  • broader native reach beyond the current browser + macOS path
  • stronger long-horizon guardian intelligence and intervention learning
  • deeper execution hardening and richer extension ergonomics
  • a fully complete end-state product; Seraph is usable now, but still under active development

Start with:

Quick Start

Recommended: Local Direct Dev Stack

# 1. Configure
cp env.dev.example .env.dev
# Edit .env.dev and set OPENROUTER_API_KEY=your-key-here

# 2. Launch
./manage.sh -e dev local up

# 3. Open
open http://localhost:3001        # Current shipped browser UI
open http://localhost:8004/docs   # Swagger API docs

# 4. Inspect / stop
./manage.sh -e dev local status
./manage.sh -e dev local logs backend
./manage.sh -e dev local down

# 5. (Optional) Screen awareness daemon
./daemon/run.sh                   # Window tracking
./daemon/run.sh --ocr             # + OCR via Apple Vision

./manage.sh -e dev local up is the canonical direct browser-development path. It explicitly loads the repo-root .env.dev, starts the backend on 8004, starts the frontend on 3001, and avoids the cwd-sensitive env drift that can otherwise change the active model/provider.

Docker Dev Stack

./manage.sh -e dev up -d
open http://localhost:3000
open http://localhost:8004/docs

Architecture

Layer Stack
Frontend React 19, Vite 6, TypeScript, Tailwind CSS, Zustand
Backend Python 3.12, FastAPI, uvicorn, smolagents, LiteLLM (OpenRouter)
Database SQLite (aiosqlite) + LanceDB (vector memory)
Tools 17 built-in tool capabilities (auto-discovered) + reusable workflows + plug-and-play MCP servers
Scheduler APScheduler — 9 jobs across briefings, reviews, strategist, observer cleanup, and memory/goal maintenance
Daemon Native macOS — window tracking, optional OCR (Apple Vision / OpenRouter)
Infra Docker Compose (backend + frontend + snekbox sandbox + http-mcp), uv

Project Structure

frontend/src/
  components/cockpit/ Guardian workspace operator surface, state rails, intervention feed
  components/        React overlays — chat, priorities panel, settings
  hooks/             useWebSocket, keyboard and operator interaction hooks
  stores/            Zustand stores — chat, priorities
  lib/               Tool parser and workspace helpers
  config/            Frontend constants

backend/src/
  api/               REST + WebSocket endpoints (chat, sessions, goals, tools, workflows, mcp)
  agent/             smolagents factory, onboarding, strategist, session manager
  tools/             @tool implementations + MCP manager
  workflows/         Reusable multi-step workflow loader, runtime, and gating
  memory/            Guardian record, LanceDB vector store, embedder, consolidator
  goals/             Hierarchical goal CRUD
  plugins/           Tool auto-discovery + registry
  scheduler/         APScheduler engine, connection manager, 9 background jobs
  observer/          Context manager, data sources, user state machine, delivery engine

daemon/              Native macOS screen daemon (window tracking + OCR)
docs/                Docusaurus docs site

Retired Surfaces

The retired village/editor line has been removed from the active repo path. Seraph is being built as a workspace-first guardian system, not a game-shell assistant.


MCP Servers

Add external tool servers with zero code changes:

./mcp.sh add things3 http://host.docker.internal:9100/mcp \
  --desc "Things3 task manager"

./mcp.sh list              # View configured servers
./mcp.sh test things3      # Test connection
./mcp.sh disable things3   # Toggle without removing
./mcp.sh remove things3    # Remove entirely

Also available via the Settings UI in the browser or the REST API (/api/mcp/servers).

Config: data/mcp-servers.json | Example: data/mcp-servers.example.json


Docker Management

./manage.sh -e dev up -d       # Start
./manage.sh -e dev down         # Stop
./manage.sh -e dev logs -f      # Tail logs
./manage.sh -e dev build        # Rebuild

Local Direct Runtime

./manage.sh -e dev local up
./manage.sh -e dev local status
./manage.sh -e dev local logs frontend
./manage.sh -e dev local logs backend
./manage.sh -e dev local down

Defaults for the local runtime:

  • frontend: http://localhost:3001
  • backend: http://localhost:8004
  • workspace dir: /tmp/seraph-dev-data
  • llm logs: /tmp/seraph-dev-logs

Override ports or local paths with env vars before launch if needed:

LOCAL_FRONTEND_PORT=3100 LOCAL_BACKEND_PORT=8100 ./manage.sh -e dev local up

Development Status

Seraph no longer uses the old phase model as the live planning surface.

Canonical docs now live in:

  • docs/implementation/ — shipped state, workstreams, and current status
  • docs/research/ — product thesis and design target

Current truth:

  • browser UI, backend APIs, observer daemon, memory, goals, and proactive scheduler foundations are shipped
  • Trust Boundaries, Execution Plane, and Runtime Reliability have strong foundations on develop
  • the source-of-truth docs now target a power-user guardian workspace and the browser app now defaults to that shell
  • no workstream is complete yet
  • Seraph still has substantial work left in presence, guardian intelligence, embodied UX, and ecosystem leverage

Start with:


Get Involved

Release History

VersionChangesUrgencyDate
v2026.7.5Seraph v2026.7.5 is the July 4, 2026 main-backed release for GPU-hosted VLM chat readiness, fast cockpit metadata, screenshot-folder reliability, local Gemma compaction, and the browser workbench journal. Release gate: - develop -> main PR #731 merged after green hosted checks on current head 4ea01269f2cbd328f7cb7cb7801ff3d04620eaf2. - PR #733 fixed the release-gate ArtifactStoragePanel test race before release. - PR #735 aligned backend/frontend/package release metadata to 2026.7.5 before releMedium7/4/2026
v2026.6.30## Summary - Releases Seraph for June 30, 2026 as v2026.6.30. - Ships the screenshot-folder source path so Seraph analyzes images written by an external screenshot producer instead of owning capture in this path. - Adds image-fact extraction/reporting and scheduled screenshot-folder ingest support. - Documents the remote VLM server path, including the companion seraph-quest/vlm-screenshot-server repo and Unsloth Gemma setup notes. ## Validation - Release PR #646 merged to main after all High6/30/2026
develop@2026-06-21Latest activity on develop branchHigh6/21/2026
v2026.4.10Latest release: v2026.4.10High4/11/2026
v2026.3.19Seraph is a local-first guardian workspace prototype with persistent memory, proactive workflows, screen awareness, and MCP integration. Highlights: - browser-based guardian workspace with activity, approvals, workflows, and operator controls - local-first backend foundations for memory, goals, scheduler, and MCP/tool execution - refreshed public repo surface, contribution docs, issue templates, and demo assets Assets: - Workspace demo MP4 - Social preview image Medium3/23/2026

Dependencies & License Audit

Loading dependencies...

Similar Packages

Micro-AgentA lightweight AI agent framework for vertical domain applications | 面向垂域应用的轻量级 AI Agent 框架v2026.08.12.3
TomoriBotA highly customizable personal AI assistant for Discord featuring smart agentic AI features such as memory, personas, tool usage, and more! | 長期記憶やペルソナ、ツール連携を完備。 次世代の「自律型AIエージェント」Discordボット!v0.7.999
dbt-mcpA MCP (Model Context Protocol) server for interacting with dbt.v2.3.0
memoraGive your AI agents persistent memory.v0.4.0
mcp-client-for-ollamaA text-based user interface (TUI) client for interacting with MCP servers using Ollama. Features include agent mode, multi-server, model switching, streaming responses, tool management, human-in-the-lv0.34.0

More in MCP Servers

difyProduction-ready platform for agentic workflow development.
tabularisA lightweight, cross-platform database client for developers. Supports MySQL, PostgreSQL and SQLite. Hackable with plugins. Built for speed, security, and aesthetics.
ai-agents-from-zero 🚀 2026 最系统的 AI Agent 速成指南|智能体实战教程 · 完整学习路径 + 实战项目 + 面试题库 · 对标大模型应用开发工程师岗位 · 覆盖LangChain / LangGraph / Coze / Dify / MCP / skills / LLM / RAG / 提示词 · 企业级部署与微调 · 从0到企业级落地 + 从学习到上线项目 + 面试准备一体化
studioOpen-source control plane for your AI agents. Connect tools, hire agents, track every token and dollar