Pramana MCP server

Pramana MCP server

mcp-name: io.github.loopg/pramana-mcp

Decide whether a prompt or model change is safe to ship, by running it against your own production runs.

Pramana is a flight recorder for AI agents. It records every non-deterministic decision your agent makes in production, replays any past run exactly against your changed code, and reports which decisions moved — not which sentences got reworded.

This MCP server lets an assistant read what Pramana recorded, and check a signed evidence bundle offline. It is read-only. There is deliberately no tool that starts a replay or a comparison: a sandboxed batch calls the model for real on every trace, so it spends money, and an MCP tool is a button any assistant can press without a person deciding. Producing a comparison stays a CLI verb behind its own confirmation gate.

Install

Nothing to install — uvx fetches it on demand.

Claude Desktop

~/Library/Application Support/Claude/claudedesktopconfig.json (macOS) or %APPDATA%\Claude\claudedesktopconfig.json (Windows):

{
  "mcpServers": {
    "pramana": {
      "command": "uvx",
      "args": ["pramana-mcp"],
      "env": {
        "PRAMANA_API_KEY": "<your-engineer-or-admin-api-key>"
      }
    }
  }
}

Claude Code

$ claude mcp add pramana --env PRAMANA_API_KEY=<your-engineer-or-admin-api-key> -- uvx pramana-mcp

Create a key in Settings → API keys at reliai.in. Use an engineer or admin key — an auditor key deliberately never receives recorded prompts or responses, so gettrace would come back empty. Self-hosting? Set PRAMANAAPI_URL to your own API.

Tools

list_traces

List recorded AI agent runs (traces) captured by Pramana, optionally filtered by agent, status or time range.

"What agent runs did we record last Tuesday?"

{ "agent": "refund-agent", "since": "2026-09-22T00:00:00Z", "limit": 20 }

get_trace

Get one recorded agent run: its steps, model calls, tool calls with arguments, and outcome.

"Walk me through what the agent did in run loan-07 — which tools did it call, and with what?"

{ "trace_id": "loan-07-5a6eef" }

Long payloads are truncated, and the response says so rather than looking complete.

listmodeldiffs

List comparison runs that checked a prompt or model change against recorded production runs.

"Have we compared anything against production since the model upgrade?"

{ "since": "2026-09-20T00:00:00Z", "limit": 20 }

getmodeldiff

Get the findings of one comparison run: which decisions changed behaviourally, which were cosmetic, and which runs halted.

"What changed when we moved to the new model last Tuesday?"

{ "diff_id": "mdr_9f2c1a", "trace_id": "loan-07-5a6eef" }

Returns bucket counts first, then root findings, then consequent ones — never merged. One root cause that produced forty downstream differences is one thing to investigate, not forty. Passing traceid is optional but much faster (see What the API could not do* below).

verify_bundle

Verify a Pramana signed evidence bundle offline, with no account and no network access.

"Here's the evidence bundle the vendor sent. Is it intact?"

{ "bundle_path": "./bundle.json", "public_key_path": "./pramana-public-key.txt" }

publickeypath is required. The public key shipped inside a bundle is never used to check that same bundle — that would defeat the point.

A pass proves the bundle has not been altered since it was signed. It does not prove that what was captured was everything that happened. This is a proof of record, not a judgement of conduct.

What the API could not do

Two tools are worse than they should be, and it is the API's fault rather than a design choice:

Shortened. The full README is on GitHub.