You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Pengfei Hu edited this page Jul 12, 2026
·
6 revisions
Agents Shipgate
The deterministic merge gate for AI-generated agent capability changes.
When a coding agent (Claude Code, Codex, Cursor) or a human changes what an AI agent can do, Agents Shipgate turns that diff into a deterministic merge verdict — before the agent gets production-like permissions. This open-source CLI and GitHub Action statically reviews the tools an agent can call: it reads a shipgate.yaml manifest, normalizes tools from MCP, OpenAPI, and SDK sources, and produces a deterministic Tool-Use Readiness Report (Markdown, JSON, SARIF).
Static by default. No agent execution. No tool calls. No LLM calls. No MCP connections. No network. No telemetry.
Sandbox controls can. Evals measure does. Agents Shipgate reviews should vs. can before merge.
Sandboxes constrain runtime capability. Evals observe behavior in tested
scenarios. Agents Shipgate compares the static tool authority introduced by a
change with the purpose, scopes, and controls reviewers intended to approve.
Static only: Agents Shipgate does not execute the agent, replace runtime
sandboxing, or prove runtime behavior.
A scan against DigitalOcean's public OpenAPI (~600 endpoints) surfaces hundreds of findings — most flagging missing approval policies on destructive operations. Against the Twilio Messaging API configured with a read-only declared purpose, it flags every Delete* operation as a scope contradiction. Against the canonical OpenAI Agents SDK customer_service example, it picks up update_seat as a write tool from name alone — without importing the file. Against the Anthropic cookbook customer service agent, it flags cancel_order as a critical missing-approval finding before the pattern reaches production.
What it is not
Not evals
Evals test behavior. Shipgate reviews release artifacts.
Not a sandbox
Sandboxes constrain runtime capability. Shipgate cannot stop a tool call in flight.
Not observability
Observability records runtime. Shipgate runs before promotion.