MCP 服务器

Tessera MCP Server

MCP server for Tessera's LLM cost-optimization layer. Anchored spend, drift detection, recommendations as typed tools.

README

tessera-mcp-server

The MCP server that returns money, not data.

Status: v0.1.2 — published 2026-05-27 on npm with sigstore SLSA provenance v1, listed on registry.modelcontextprotocol.io.

Tessera is an LLM proxy that optimizes API spend through multi-provider routing, prompt compression, audit-immutable logging, output-length prediction, and batch arbitrage. This package exposes Tessera as an MCP server for tool-using agents — Claude Desktop, Claude Code, Cursor, Cline, Continue, Goose, Zed.

Where other LLM-infrastructure MCP servers return logs, traces, or prompt metadata, this one returns savings decisions: what's drifting, what to switch, what to approve, what to audit.

Install

npx @tessera-llm/mcp-server

Or add to your client config:

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json on macOS, %APPDATA%\Claude\claude_desktop_config.json on Windows):

{
  "mcpServers": {
    "tessera": {
      "command": "npx",
      "args": ["-y", "@tessera-llm/mcp-server"],
      "env": {
        "TESSERA_API_KEY": "tk_..."
      }
    }
  }
}

Claude Code (.mcp.json in project root):

{
  "mcpServers": {
    "tessera": {
      "command": "npx",
      "args": ["-y", "@tessera-llm/mcp-server"],
      "env": {
        "TESSERA_API_KEY": "tk_..."
      }
    }
  }
}

Cursor (~/.cursor/mcp.json):

{
  "mcpServers": {
    "tessera": {
      "command": "npx",
      "args": ["-y", "@tessera-llm/mcp-server"],
      "env": { "TESSERA_API_KEY": "tk_..." }
    }
  }
}

Get a TESSERA_API_KEY at tesseraai.io/dev — Free Sandbox is 60M tokens/month with no card.

Tools

v0.1 exposes 6 tools (5 read + 1 mutate). Hard cap — no tool sprawl.

Tool	Read/Write	Purpose
`tessera_list_workloads`	read	List your mapped workloads with anchor cost + current m-stack.
`tessera_get_savings_report`	read	Anchored spend + measured savings + performance-fee accrual for a window.
`tessera_get_recommendation_queue`	read	Pending Optimize-tab recommendations with expected lift + confidence.
`tessera_get_ledger_entries`	read	Audit-immutable Monthly Reading rows (provider call, mechanic stack applied, savings).
`tessera_get_quality_snapshot`	read	SLA floor + p50/p95 quality scores + drift events.
`tessera_approve_recommendation`	mutate	Move a queued mechanic from "suggested" to "active" with audit-trail entry.

Provider config writes, API-key management, composition cap changes, and Stripe operations are deliberately NOT in this surface — they live in the dashboard, where blast-radius requires explicit modal confirmation.

Transport

stdio (default) — local clients (Claude Desktop, Cursor, Cline, Continue, Claude Code)
Streamable HTTP (optional) — set TESSERA_MCP_TRANSPORT=http to bind on localhost:8788 for remote / Goose / Zed-via-mcp-remote

SSE (deprecated in MCP spec 2025-11-25) is not supported.

Auth

TESSERA_API_KEY env var or Authorization: Bearer <key> header (HTTP transport). Same API key as the SDK — tk_* format. Future v0.2: OAuth 2.1 (aligned with MCP spec RC 2026-07-28).

Security posture

Tools receiving user-controlled content (ledger notes, recommendation rationale strings, workload names) are labelled __untrusted__ to prevent prompt-injection cascade per the Supabase/Cursor 2025 pattern.
mcp-scan (Invariant Labs) runs in CI to catch tool-poisoning attacks in tool descriptions.
No execute_code escape hatch. Typed verbs only.
No session-based auth. Every request authenticates independently per MCP spec 2026 requirement.