Put a signed record on every MCP tool call.
dh proxy sits between your agent's MCP client and the MCP server it uses. The agent sees the same tool list, byte for byte. Every consequential call gets a policy decision and a signed record before it reaches the server.
The gap it closes
Most agents reach real systems through MCP servers: file systems, databases, ticketing, cloud APIs. The server sees a call. It doesn't see who sponsored the agent, what the agent had read, or whether this call should run at all.
Capabilities
No change to the agent
Point the agent's MCP config at dh proxy instead of the server. Same tools, same schemas, same results.
stdio and HTTP
Run a local server as a child process, or govern a hosted MCP server over streamable HTTP without touching it.
Decides before the call
Policy runs outside the model. Blocked calls never reach the server; checkpoints wait for an approver.
Knows what the agent read
Reads, resources and prompts are recorded as the agent's horizon, with content hashes, so a later write can be judged against them.
States what it can't see
Every record names the proxy as its insertion point and lists what that point can't observe.
One chain per session
Each session is a hash-chained file. SIGTERM writes the closing record before the process exits.
dh keygen
dh init
dh proxy -c dh.toml -- npx -y @modelcontextprotocol/server-filesystem /path/to/repo
# hosted server, governed without touching it
dh proxy -c dh.toml --listen 127.0.0.1:8787 --downstream-url https://tools.example.com/mcpWhat it doesn't do
- A call that doesn't pass through the proxy leaves no record, so keep the server reachable only through it (a token only the proxy holds, or a private network).
- One proxy process is one record chain; run one per agent session.
See it on one of your own workflows.
We'll put Decision Horizon in front of a real agent workflow of yours and replay the record with you.