Guardrails
Workspace policies that control tool access and sanitize outputs before they reach the model.
Source: https://runstack.engineer/docs/guardrails
Guardrails are workspace-scoped policies that control what tool outputs reach the model. Credentials stay vaulted on Runstack — these policies apply to every chat in the workspace.
Configure them under Settings → Workspace → Guardrails.
Output sanitization
| Setting | Default | Description |
|---|---|---|
| Secret redaction | On | Redacts API keys, tokens, JWTs, Bearer headers, and key=value secrets from tool outputs |
| PII scrubbing | On | Redacts emails, US phone numbers, and SSN-like patterns from tool outputs |
Object keys matching sensitive names (token, secret, password, api_key, authorization, credential, auth) are replaced with [REDACTED_SECRET].
Tool policies
| Setting | Default | Description |
|---|---|---|
| Blocked tools | Empty | One tool name per line — blocked before execution |
| Allowed tools (optional) | Empty | Allowlist; when set, only listed tools can run. Empty means all non-blocked tools are allowed |
| Max tool calls per turn | 50 | Integer from 1 to 200 — caps how many tool calls the agent can make in a single turn |
Click Save changes to apply. Guardrail instructions are also injected into the agent system prompt.
Enforcement
When a policy blocks execution, the agent receives an error such as:
- Tool blocked by workspace guardrails
- Tool call limit reached for this turn
Output sanitization runs on MCP and meta-tool results before they are returned to the model.
Guardrails apply to workspace chat and Studio chat. They also apply to MCP execute_tool calls authenticated with workspace API keys.