Wallras clients for Python

Two single-file, zero-dependency clients for Wallras, the AI monitoring proxy you run next to your app:

• wallras_mcp.py: an MCP server, so Claude Desktop, Claude Code, Cursor or your own agent can check drift, read requests and agent runs, work the review queue, fetch prompts and run regression sets. Read-only unless you pass --enable-writes.
• wallras_sdk.py: report what Wallras can't see on its own: a self-hosted model's token probabilities and decision records, agent steps (tool runs, lookups), and feedback; fetch prompt versions.

Both talk only to your own Wallras with your token. The Wallras engine itself ships as a Docker image (ghcr.io/wallras/wallras-engine); this repository holds the open-source clients.

Install

pip install git+https://github.com/wallras/wallras-python

Or run the MCP server without installing anything:

uvx --from git+https://github.com/wallras/wallras-python wallras-mcp

Each file also works on its own: copy it next to your code.

MCP server

{"mcpServers": {"wallras": {
  "command": "uvx", "args": ["--from", "git+https://github.com/wallras/wallras-python", "wallras-mcp"],
  "env": {"WALLRAS_URL": "http://localhost:8080", "WALLRAS_TOKEN": "a viewer or admin token"}}}}

Add "--enable-writes" to args for the tools that change something (feedback, review decisions, saving or promoting prompts, regression runs, optionally graded by a judge_model). Those need an admin token; regression runs call your provider with WALLRAS_PROVIDER_KEY (and WALLRAS_ANTHROPIC_KEY / WALLRAS_GEMINI_KEY), which stay on your machine.

SDK

from wallras_sdk import Wallras
w = Wallras("http://localhost:8080", token="your WALLRAS_APP_TOKEN")
w.report(model="llama-3.1-8b", prompt=prompt, answer=text,
         tokens=Wallras.tokens_from_openai(response["choices"][0]["logprobs"]["content"]))

Reports go in the background and never raise into your code. See the docs for traces, decision records and prompts.

Try a hub with two nodes

examples/docker-compose.hub.yml starts one hub and two nodes on your machine:

docker compose -f examples/docker-compose.hub.yml up -d

Send requests to a node (http://localhost:8081/v1 or :8082, with your provider key), then open http://localhost:8080/wallras/ui and the Nodes tab.

Try active-active

examples/docker-compose.active.yml starts Postgres, three Wallras containers and a load balancer. Any container answers any request, and together they keep one tamper-evident record:

docker compose -f examples/docker-compose.active.yml up -d

Send requests to http://localhost:8080/v1, then ask the MCP server's wallras_fleet tool (or GET /wallras/fleet) which containers answered.

Issues

Bugs and ideas for the clients, or for Wallras itself: open an issue. Security reports: see SECURITY.md, not public issues.

MIT licensed. The files here are copies of the ones published with each Wallras release.