OpenAI compatibility

Compare Rebyte with the managed OpenAI Agents API and check current limitations.

Agents API5 min read

On this page

Available capabilities Features not yet implemented Behavior and protocol differences Rebyte-specific extensions

Rebyte implements a subset of the managed OpenAI Agents API. Use the official openai client with baseURL: 'https://api.rebyte.ai/v1' and your Rebyte key. Using that client does not mean every OpenAI endpoint or option is implemented.

This comparison was reviewed on September 25, 2026 against OpenAI's current documentation and the deployed Rebyte implementation. History retrieval was also tested against production with openai@7.15.0: HTTP 200, saved user and assistant messages, pagination, and identical results from a fresh client. The remaining rows describe implementation review, not a new end-to-end test of every capability.

The scope is the managed Agents API. The application-side @openai/agents runtime, Responses API and Conversations API are separate integrations; their features are not automatically Agents API compatibility requirements.

Available capabilities

Capability	Rebyte
Saved Agents and per-Session configuration	Create, list, retrieve, update and delete Agents; create Sessions from saved or inline configuration. Existing Sessions retain their configuration snapshot.
Conversation history	List saved Items, including messages and tool output; retrieve Turns and continue the same Session.
Live output and cancellation	Native Session/Turn/Item events and text deltas. Cancel active work while retaining the Session. Both services require saved Items to recover missed events.
Application functions	`required_actions`, function results, failure results and deferred function discovery with `tool_search`. Image results are excluded; see the limits below.
MCP	HTTP and stdio connections, service or environment execution where supported, tool discovery, resources and Vault credentials.
Managed compute and files	Rebyte-managed Session environments, command execution, patches, images, inline file uploads and immutable Artifacts.
Basic web search	Live search, domain filters and result context size.
Skills	Inline ZIPs, Rebyte's GitHub-sourced Skills, and capability directories.
Execution tracing	Rebyte Platform trace views through Langfuse; public Sessions, Turns and Items for inspecting saved outcomes.

See OpenAI's Session guide and Rebyte's overview for the common execution flow.

Features not yet implemented

OpenAI Agents API capability	Current Rebyte gap
Automatic context compaction	No automatic summarization/compaction of long Session history. Persisting history does not remove the model's context-window limit.
Multi-agent orchestration	`multi_agent.enabled: true` is rejected. No compatible subagent lifecycle or subagent history endpoints.
Session Webhooks	No signed Agents API lifecycle callbacks. Applications must use streaming and/or retrieve Session state. Legacy Task webhooks do not provide this contract.
Self-hosted environments	No `environment.type: 'self_hosted'` executor connection flow. Choose Rebyte-managed compute or `none`.
Environment templates	No public Environment Template CRUD or `environment_template_id` selection. Rebyte's internal VM templates are a different resource.
Plugins	Nonempty `environment.plugins` is rejected. Installing a Skill is not equivalent to loading a plugin's manifest and MCP configuration.
Updating an existing Session's model/settings	Session update currently accepts `metadata` only. It cannot update `agent.model`, `agent.reasoning.effort`, or `agent.service_tier` while keeping the same Session.
Public OTLP trace export	No compatible `GET /v1/agents/sessions/{session_id}/traces`. Rebyte's Platform trace view does not implement this public export endpoint. OpenAI requires trace export to be enabled for the organization.
Files API ID inputs	`environment.files` accepts inline base64 files, not `type: 'file_id'` references to OpenAI's Files API.
Vault secrets for sandbox HTTP requests	No `environment_variable` Vault credential or network-proxy placeholder substitution. Rebyte Vaults support service-origin MCP bearer/OAuth credentials. Ordinary environment variables expose their values to sandbox code.

Behavior and protocol differences

Area	OpenAI Agents API	Rebyte
Input during active work	A new message can steer the active Turn.	A new message creates a queued Turn and waits for the active Turn. See input queuing.
Programmatic tool calling	`programmatic_tool_calling`, enabled by default; generated JavaScript can invoke application functions through their normal handoff.	Opt-in `dynamic_workflow` exposes `run_code`, with a 300-second limit and server-executable tools only. Different declaration, runtime contract and function boundary.
Web search options	Supports live/cached/disabled modes and location targeting.	Cached mode and non-null location return 400. Live and disabled modes are supported.
Function results	Supports strings and supported content arrays.	Text results work; application function image results are explicitly rejected. This does not mean user image input or the environment's `view_image` tool is unsupported.
Skill discovery	The harness discovers `SKILL.md` files in capability directories and adds their metadata to context.	Inline/GitHub Skills are installed; configured capability directories are described to the model for inspection with environment tools. No equivalent automatic directory catalog scan in the managed loop.
Model controls	Available settings depend on OpenAI's model.	Model-specific provider adapters determine reasoning, verbosity and JSON Schema support; `service_tier: 'fast'` is rejected.

Sources: OpenAI Session input, Programmatic Tool Calling, Web search, Functions, and Agents API Skills. For Rebyte details, see Dynamic Workflow, Functions and Agent settings.

Rebyte-specific extensions

Workflow Agents, Scheduled Agents, Platform connections, GitHub Skill sources and Dynamic Workflow extend Rebyte's API. They are not evidence that the corresponding OpenAI tools, templates or multi-agent protocol are implemented. Workflow and Schedule resources use the optional @rebyteai/agent-extensions package; ordinary Agent calls use openai directly.

For app integration, AppKit restores saved history but does not automatically execute application functions. That client-side handler is separate from the backend's existing function-call support.