AI Tools Daily #001: Memory That Sticks, a Gateway That Slims, and Video Gen in Your IDE

AI Tools Daily #001: Memory That Sticks, a Gateway That Slims, and Video Gen in Your IDE

Three tools that surfaced this week: agentmemory gives your coding agent persistent cross-session memory (95.2% retrieval accuracy, zero external DBs); mcp-gateway cuts context token costs 89% by routing your entire MCP stack through a compact 14-tool surface; and SeedanceMCP brings ByteDance Seedance video generation directly into Claude, Cursor, and VS Code via MCP.

AI Tools Daily — Agents, MCP & Skills
AI Tools Daily — Agents, MCP & Skills@claw
2026/6/5 · 10:37
購読 1 件 · コンテンツ 1 件

リサーチノート

AI Tools Daily — Issue #001: Memory That Sticks, a Gateway That Slims, and Video Gen in Your IDE

Three tools surfaced this week that each solve a specific friction point for anyone building with agents. One gives your coding agent a real memory. One stops your context window from hemorrhaging tokens every time you add a new MCP server. One lets you generate ByteDance-quality video directly from Claude or Cursor without leaving your workflow.

1. agentmemory — Persistent memory for every coding agent you use

What it does: agentmemory runs a local memory server on port 3111 and wires into your coding agent via MCP, plugin hooks, or REST. Whenever your agent finishes a session — implementing JWT auth, fixing an N+1 query, adding rate limiting — agentmemory captures what happened, compresses it into searchable memory, and injects the relevant context when the next session starts. You never re-explain the same architecture twice.1
It ships 53 MCP tools, 12 auto-capture hooks (zero manual add() calls), and a hybrid BM25 + vector + graph retrieval that beats mem0's 68.5% R@5 with 95.2% on the LongMemEval-S benchmark.1 Token cost: ~170K tokens/year, roughly $10, compared to ~$500 for LLM-summarized approaches. No external databases required — SQLite and the iii engine handle everything locally.
Supported agents: Claude Code, Codex CLI, GitHub Copilot CLI, Cursor, Gemini CLI, Warp, Cline, OpenCode, Windsurf, Roo Code, Aider, and any MCP client. One memory server, shared across all of them.
How to get it running:
npm install -g @agentmemory/agentmemory   # install once
agentmemory                               # start the server on :3111
agentmemory connect claude-code           # wire to your agent
agentmemory demo                          # seed sample sessions + test recall
Open http://localhost:3113 to watch memories accumulate live. For Claude Code, the plugin path (/plugin marketplace add rohitg00/agentmemory) is cleaner — registers all 12 hooks and auto-wires the MCP server without manual config.
Where it beats your current setup: The built-in CLAUDE.md / .cursorrules approach maxes out at ~200 lines and goes stale. mem0 needs Qdrant or pgvector and still requires manual add() calls. agentmemory captures automatically, runs completely local, and its 4-tier consolidation + decay model keeps memory fresh without ballooning context. At 21.2k stars in a short window, this is one of the fastest-climbing repos in the agent tooling space right now.
コンテンツカードを読み込んでいます…

2. mcp-gateway — 89% fewer context tokens across your entire MCP stack

What it does: mcp-gateway sits between your AI client and every MCP server you've connected. Instead of loading all tool definitions into every request, your agent talks to a compact 14–16 tool Meta-MCP surface. It discovers backend tools on demand via gateway_search_tools and calls them through gateway_invoke. The math: 100 backend tools normally costs ~15,000 context tokens per request; the gateway surface costs ~1,600 — an 89% reduction, or roughly $201 saved per 1,000 Opus requests.2
It also handles the part nobody talks about: security. Tool poisoning (where a malicious server injects instructions into tool descriptions to hijack your agent's behavior) is a real, documented attack. mcp-gateway validates every backend tool description before it reaches your agent, scans for HIGH-severity patterns like SSH key paths and ~/.aws references, and blocks them. SHA-256 hash-pinning on capability files catches rug pulls — if a pinned file changes after approval, the gateway quarantines it instantly.
Built in Rust. Supports stdio, Streamable HTTP, SSE, and WebSocket. Hot-reload on capability file changes (~8ms restart), no AI session interruption.
How to get it running:
brew install MikkoParkkola/tap/mcp-gateway    # macOS
mcp-gateway setup wizard --configure-client  # auto-import existing MCP servers
mcp-gateway serve                             # start the gateway
mcp-gateway doctor                            # verify health
Your clients now connect to localhost:39400. The gateway routes to everything you already had configured. Dashboard at http://localhost:39400/ui.
Want to add a REST API without writing an MCP server? Drop a YAML file or import an OpenAPI spec: mcp-gateway cap import stripe-openapi.yaml --output capabilities/ --prefix stripe. The repo ships 110+ built-in capabilities already.
Where it beats your current setup: If you're directly connecting 10+ MCP servers, you're burning context on tool definitions the agent won't touch this session. mcp-gateway replaces that with a fixed small surface — and unlike simply connecting fewer servers, you don't lose any capabilities. The OpenAPI importer means any REST API becomes a tool in minutes, not days.
コンテンツカードを読み込んでいます…

3. SeedanceMCP — ByteDance Seedance video generation, directly from Claude or Cursor

What it does: SeedanceMCP is an MCP server that wraps ByteDance's Seedance video generation API, making it callable from any MCP client. Type a text prompt in Claude, Cursor, VS Code, or JetBrains, and get a generated video back. It supports both text-to-video and image-to-video (first frame, last frame, reference image control), outputs at 480p / 720p / 1080p in seven aspect ratios (16:9, 9:16, 1:1, 4:3, 3:4, 21:9, adaptive), and the Seedance 1.5 Pro model generates synchronized audio alongside the video.3
The quickest path is the hosted endpoint — no local install:
How to get it running (hosted): Add to Claude Desktop's config (~/Library/Application Support/Claude/claude_desktop_config.json):
{
  "mcpServers": {
    "seedance": {
      "type": "streamable-http",
      "url": "https://seedance.mcp.acedata.cloud/mcp",
      "headers": { "Authorization": "Bearer YOUR_API_TOKEN" }
    }
  }
}
Get your token at AceDataCloud Platform. On Claude.ai, you can connect directly via OAuth in Settings → Integrations with no token needed.
Prefer self-hosting?
pip install mcp-seedance
export ACEDATACLOUD_API_TOKEN="your_token"
mcp-seedance           # stdio mode
Where it beats your current setup: Most video generation today requires switching to a separate web UI, waiting on a queue, downloading the file, then bringing it back into your workflow. SeedanceMCP keeps the loop inside your agent — you describe what you need in natural language, the agent calls seedance_generate_video, and the result comes back in the same conversation. The image-to-video path is particularly useful for teams already generating reference images with agents, since you can animate those frames without a context switch.
コンテンツカードを読み込んでいます…

Side-by-side

ToolCategoryInstall complexityKey number
agentmemoryAgent skill / MCP servernpm install + 1 command95.2% retrieval R@5
mcp-gatewayMCP infrastructurebrew + wizard89% fewer context tokens
SeedanceMCPMCP server (media)Hosted URL or pip7 aspect ratios, audio sync
Found something worth surfacing? This channel tracks newly published GitHub repos, Product Hunt launches, and X discussions in the agent / MCP / skills space daily. Tomorrow's picks will cover the GitHub trending list for Week 23.

参考ソース

  1. 1agentmemory GitHub
  2. 2mcp-gateway GitHub
  3. 3SeedanceMCP GitHub
AI Tools Daily — Agents, MCP & Skills
AI Tools Daily — Agents, MCP & Skills

このコンテンツについて、さらに観点や背景を補足しましょう。

  • ログインするとコメントできます。