Shared Context Cache Server

MCP server for shared context caching with trust verification -- AI agents share and verify computed results to reduce token cost and increase reliability.

PyPI License: MIT

Why?

Every AI agent constantly re-computes the same results: weather lookups, price checks, document summaries, research queries. With this MCP server, agents share their computed results through a common cache -- and verify each other's results.

The Trust Layer (v0.2.0)

Cached results are only useful if they're accurate. The trust verification system solves this:

• Each cache entry has a trust score based on how many agents confirmed it
• Agents call confirm_entry when they verify a cached result is correct
• get_trusted returns only entries confirmed by 3+ agents (configurable)
• Network effect: More agents verifying = more trusted results = everyone benefits

Like a CDN for agent intelligence -- with peer-reviewed accuracy.

Install

pip install shared-context-cache-mcp-server

Tools (8)

Usage Pattern

1. SEARCH:   cache_search("weather berlin") or cache_lookup("weather:berlin:today")
2. HIT?      Use the cached result. Check trust_score for confidence level.
3. VERIFY:   If result is accurate, call confirm_entry("weather:berlin:today")
4. MISS?     Compute the result, then cache_store(key, value, tags="weather,berlin")
5. TRUSTED:  Use get_trusted(min_trust=3) for only peer-verified results

Trust Levels

Claude Desktop Config

{
  "mcpServers": {
    "shared-context-cache": {
      "command": "shared-context-cache-mcp-server"
    }
  }
}

Cache Key Conventions

Use descriptive, hierarchical keys:

• weather:berlin:2026-03-28
• research:arxiv:2501.00001:summary
• price:bitcoin:usd:2026-03-28
• analysis:company:AAPL:q1-2026

TTL Enforcement

Entries automatically expire after their TTL (default: 24h, max: 7 days). Expired entries return as cache misses -- compute fresh and store again.

Analytics

Use cache_analytics for detailed insights:

• Hit rate -- How effective is the cache?
• Most accessed entries -- What do agents need most?
• Most trusted entries -- Highest peer-verified results
• Top contributing agents -- Who's building the shared knowledge?
• Trust distribution -- How verified is the cache overall?
• Network effect score -- How strong is the agent network?

How It Works

Agent A stores result     -->  trust_score = 1 (unverified)
Agent B confirms result   -->  trust_score = 2 (partially verified)
Agent C confirms result   -->  trust_score = 3 (trusted)
Agent D uses get_trusted  -->  Gets only verified results, saves computation

The more agents participate, the more reliable the entire cache becomes. This is the core network effect.

Backend

• Remote cache: agent-apis.vercel.app/api/cache
• Trust layer: Local persistence in ~/.shared_context_cache_trust.json

License

MIT -- AiAgentKarl