ProvelyOpenAI Agents SDK connects to Provely through the MCP server, the CLI, the SDK, the REST API. The agent does the action. Provely verifies the outcome and signs a receipt. OpenAI Agents SDK never holds the verifier credentials.
A person verified the OpenAI Agents SDK capabilities on 2026-09-08. Agent products change monthly. Re-verify before you build.
Source: integration profile openai-agents · retrieved 2026-09-08
Which surfaces does OpenAI Agents SDK use?
| Surface | Support | Detail |
|---|---|---|
| MCP server | Supported | transport stdio, http |
| CLI | Supported | Runs through the shell tool of the agent. |
| SDK | Supported | languages python, typescript; two lines, no adapter: OpenAI Agents |
| REST API | Supported | Any HTTP call reaches the control plane. |
Which guidance files does OpenAI Agents SDK read?
Provely ships one Agent Skill for this host, so the agent knows to verify an outcome before it claims done. The renderer writes prompt fragment, tool schemas from one source. A new agent needs a profile, not core code.
Explain: each file, its path and its write mode
| Format | Path | Mode |
|---|---|---|
| prompt fragment | provely/prompt-fragment.md | fragment |
| tool schemas | provely/tool-schemas.json | fragment |
| CI gate workflow | .provely/ci/github-actions.yml | fragment |
How is a false completion claim blocked?
Guidance asks. A hook and the CI gate enforce. A host with neither still receives the verdict, and a person reads it before the work ships.
Explain: what each control does on this host
| Enforcement | Support | Detail |
|---|---|---|
| Advisory guidance | Supported | The instruction files state the rules for every agent. |
| Lifecycle hooks | Supported | tool_input_guardrail, on_invoke_tool, output_guardrail |
| CI gate | Supported | provely verify --wait fails the pipeline unless the verdict is VERIFIED. |
How do I connect OpenAI Agents SDK?
Connect OpenAI Agents SDK to Provely
- Install the CLI.Run
curl -fsSL https://provely.sh/install.sh | sh. On Windows, runirm https://provely.sh/install.ps1 | iex. Each script verifies the signed checksum before it installs the binary. - Register the MCP server.Add the Provely server. Use
provely mcpfor stdio, or the hosted HTTP server with a bearer token. The tools arebegin,verify,status, andreceipt. - Write the guidance files.Run
provely init --agent openai-agents. It writesprovely/prompt-fragment.md(fragment),provely/tool-schemas.json(fragment),.provely/ci/github-actions.yml(fragment). The guidance carries one rule: a successful tool call is not completion. - Install the hooks.Enable the
tool_input_guardrail,on_invoke_tool,output_guardrailhooks. A hook calls the CLI and blocks a completion claim while an operation is PENDING. - Add the CI gate.Run
provely verify --wait <duration>in the pipeline. The step fails unless the verdict is VERIFIED. Exit codes: 0 VERIFIED, 2 PENDING, 3 FAILED, 4 CONTRADICTED, 5 UNVERIFIABLE, 1 error. - Report the verdict exactly as returned.VERIFIED: "The action is verified complete. Receipt: <id>." PENDING: "The action is accepted but not yet verified. Operation: <id>."
What else does the profile state?
- Corrected 2026-09-08: an output guardrail blocks the final answer only. It does not stop a tool call. The earlier profile named it as the one enforcement point, and that overstated the cover.
- Verified 2026-09-08: @tool_input_guardrail runs before a function tool runs. reject_content blocks the call, and raise_exception stops the run. A blocking @input_guardrail stops the whole run. RunHooks and AgentHooks only observe. Source: https://openai.github.io/openai-agents-python/guardrails/.
- Verified 2026-09-08: this repository wraps FunctionTool.on_invoke_tool and reads the ask through AgentHooks.on_llm_start. Read sdks/python/src/provely/guard/openai_agent.py and the TypeScript twin. The block happens at the tool boundary.
- The interception point is code, not a script. It blocks only after the application calls instrument. An application that never calls it gets advisory guidance and the CI gate.
- Verified 2026-09-08: the SDK reads MCP servers through MCPServerStdio, MCPServerStreamableHttp, and MCPServerSse. A remote server takes headers with Authorization inside params. Source: https://openai.github.io/openai-agents-python/mcp/.
- Two version lines exist. The profile states one range that covers both. The adapter records the SDK version in effect per operation.
OpenAI Agents SDK keeps its action credentials, and no surface returns the verifier credentials to it. Read the credential boundary. An operation outlives the session that opened it. Read the API reference.
Which host versions does the profile cover?
>=0.17.0. The integrations team last verified the profile with openai-agents 0.22.1 of 2026-09-08 and @openai/agents 0.17.2 on 2026-09-08.