Frameworks

Pick how Faramesh wires into your agent. SDK shim, MCP proxy, HTTP proxy, or A2A.

Faramesh intercepts tool calls at the layer that best fits your agent runtime. There are four interception tiers, and most agents only need one.

Tier	Mechanism	When to use it	Frameworks
1. SDK shim	One-line wrapper around your tool list.	Native code agents you control.	LangGraph, LangChain, CrewAI, OpenAI Agents, Google ADK
2. MCP proxy	The agent talks MCP; Faramesh sits in the middle.	Off-the-shelf MCP clients you can't modify.	Claude Code, Cursor, OpenCode
3. HTTP proxy	Faramesh sits in front of the model/tool HTTP endpoint.	Vendor-hosted agent runtimes.	Bedrock, AgentOS
4. A2A proxy	Cross-agent delegation gateway.	Multi-agent and cross-vendor setups.	Any A2A-speaking agent

Pick the tier that matches your runtime, then read the per-framework page.

Native SDK frameworks

These get the SDK shim. You import GovernedToolSet (Python) or governedTools (TypeScript) and pass it your tool list. Every tool call routes through the Faramesh daemon before execution.

MCP-based agents

Point the MCP client at the Faramesh proxy URL. Faramesh forwards approved calls to the real MCP server and denies / defers everything else.

Hosted runtimes

Faramesh terminates HTTP at the proxy port and forwards approved calls to the vendor endpoint.

Bedrock

How tiers compare

Property	SDK shim	MCP proxy	HTTP proxy
Latency overhead	< 2 ms	3–8 ms	5–15 ms
Code change	One line	Zero (config only)	Zero (config only)
Tool-argument visibility	Full	Full	Full
Streaming responses	✓	✓	partial
Bypassable from inside the process	Yes (mitigated by OS-tier on Linux)	No	No

For untrusted agents, combine SDK-tier with Linux OS-tier enforcement (enforcement { os_tier = true }).

What's next

Stack reference: enforcement block per agent
Quickstart: end-to-end walkthrough with LangGraph
CLI: operating Faramesh once your framework is wired