witnora

Guided Assurance Copilot v0.1

The Hosted Setup page turns a plain-language description of one agent workflow into a typed agentcert.setup_plan.v0.1 plan. The default planner is deterministic and does not require a model API key.

Flow

1. Describe one agent and its highest-risk action.
2. Witnora classifies the nearest boundary: browser, coding, MCP, workflow, or data.
3. Review one recommended next action and its evidence limitation.
4. Explicitly authorize credential creation, integration generation, a test run, or a GitHub pull request.

5. Download the redacted plan and apply it inside the customer repository:

   npx witnora@latest setup apply --plan witnora-setup-plan.json --repo .
   

   Add --github-pr only after reviewing the plan. The customer-side runner creates a branch and pull request through authenticated git and gh. --configure-ci may be added when WITNORA_PROJECT_ID and WITNORA_API_KEY are already present in the local environment. The ID is stored as a GitHub Variable and the key is passed to gh secret set over standard input.

6. Upload evidence and inspect it in the Hosted split-pane viewer. JSON, JSONL, images, HTML, and PDF are previewed without a local download; ZIP and unsupported formats remain metadata-only downloads.

Authorization boundary

The planner cannot execute commands, create credentials, modify a repository, or run tests. Authorization is recorded separately from execution. Plans never contain plaintext API keys, and the runner ignores commands embedded in a plan: it can only generate Witnora-owned allowlisted templates.

Browser and MCP templates can become release evidence after their deterministic Tripwire or MCPBench artifacts exist. Coding, workflow, and data templates initially record an observed event only. They do not establish CURRENT until the customer adds a deterministic task assertion and outcome verifier.

Optional model interpretation

An operator may configure an OpenAI-compatible interpreter with:

• WITNORA_SETUP_INTERPRETER_API_KEY
• WITNORA_SETUP_INTERPRETER_MODEL
• WITNORA_SETUP_INTERPRETER_URL (optional)
• WITNORA_SETUP_INTERPRETER_TIMEOUT_MS (optional, default 10 seconds)

The model can only propose the typed intent fields. Invalid, timed-out, or failed responses fall back to the deterministic classifier. Model output cannot add commands, credentials, operation kinds, or executable steps.

What this proves

• The user authorized each sensitive setup operation.
• The generated integration came from a versioned, allowlisted template.
• Hosted received and retained the first evidence object.
• The conversation-to-evidence duration is measurable.

It does not prove an agent is reliable merely because setup completed. That requires relevant evidence, a declared scope, and an applicable assurance decision.

This site is open source. Improve this page.