Governed model gateway

Wardwright

Agents call normal OpenAI-compatible model names. Wardwright decides what those names mean: which provider or local model to use, what policy rules apply, when to retry, reroute, block, or rewrite output, and what receipt should be recorded afterward.

Install Policy Workbench Model Middleware Agent Authoring Agent Adapters Framework Adapters Tool Mediation News Monitor Tutorial GitHub
Status: Wardwright is early but installable. The v0.0.11 release line includes the Lustre workbench migration, framework-adapter recipes, local agent-adapter install/probe support, and the first Wardwright-hosted server-tool plus tool-mediation controls.

Wardwright is for teams and operators who want model behavior to be a reviewed, testable contract instead of scattered prompt strings, provider IDs, and retry logic inside every agent.

Today, Wardwright can run as a local or remote service, define Wardwright models, simulate policy behavior in the /admin workbench, record receipts, and exercise early policy examples such as routing decisions, stream governance, output checks, retries, and saved simulator test cases. Operator workflows start from /admin; old /policies links redirect there. The admin surface currently supports basic auth, while individual models can be configured for API-key or open access.

Install

Wardwright publishes native binaries through GitHub Releases and Homebrew.

macOS Homebrew

brew tap bglusman/tap
brew install wardwright
wardwright admin

Linux

curl -fsSL https://raw.githubusercontent.com/bglusman/wardwright/main/scripts/install.sh | sh
WARDWRIGHT_SECRET_KEY_BASE="$(openssl rand -base64 64)" \
WARDWRIGHT_BIND=127.0.0.1:8787 \
~/.local/bin/wardwright serve

Set WARDWRIGHT_ADMIN_TOKEN before exposing Wardwright beyond loopback. See Packaging for manual archive install steps and service details.

For an end-to-end example of why the workbench and framework recipes matter, see Build A Lightweight News Monitor.

What Wardwright Adds

Wardwright is a narrow control point between agent code and concrete model providers.

Stable Model IDs

Expose names such as coding-balanced while Wardwright owns the model graph and policy behind them.

Model Composition

Route to provider targets, delegate through other Wardwright models, and keep context-window and fallback choices explicit.

Policy And Repair

Apply request, route, stream, output, history, alert, and tool controls, including bounded stream retries and rewrites.

Receipts

Record caller provenance, selected routes, provider attempts, policy actions, and final status for each run.

Policy Workbench

The installed app includes a workbench at /admin for visualizing and simulating Wardwright models before using them behind real traffic. Choose the registered Wardwright model being simulated, load a fixture, edit the simulated request, model output, and retry attempts, then step through the resulting policy run.

Wardwright registered-model workbench showing a retry simulation fixture
The simulator can replay stream governance, including editable model output, retry attempts, state transitions, and receipt evidence.

Fresh installs include starter examples for output contracts, route/model composition, stream repair and session state, plus tool/workflow control. Locally authored models use the same workbench path when they expose a supported projection, and reviewed turns can be saved as reusable simulator test cases.

See Policy Workbench for screenshots and examples. External agents can use wardwright tools, /mcp, and the protected authoring APIs; see Agent Authoring for the review workflow. Adapter install and validation commands are covered in Agent Adapters. SDK and application-framework recipes are tracked separately in Framework Adapters.

Provider Credentials

Wardwright can call local Ollama without credentials. OpenAI-compatible provider targets can reference secrets through credential_fnox_key or credential_env. Fnox is a secret lookup path, not Wardwright authentication; keep real provider credentials on loopback-only instances or behind a trusted auth boundary. See Provider Credentials.

Current Runtime

The active app is a Phoenix service with a Lustre operator workbench. Elixir owns Phoenix, HTTP/API boundaries, provider calls, storage drivers, PubSub, and top-level supervision; Gleam owns correctness-heavy policy logic and new workbench behavior. Runtime state is being evaluated for gleam_otp migration where typed actors can remove invalid states or clarify concurrency ownership.

Current capabilities include: