System messages. OpenAI puts the system instruction as a regular message with role: "system" . Anthropic puts it at the top level as a system field on the request. We move it across and concatenate if you have multiple. (Anthropic accepts a single string for system; multiple OpenAI system messages get joined with double newlines.) Tool calls (assistant side). OpenAI: tool_calls: [{id, type:'function', function:{name, arguments:'JSON-string'}}] . Anthropic: a tool_use block inside the assistant's content array with {id, name, input} (parsed object, not a string). We parse OpenAI's escaped JSON arguments into the structured input Anthropic expects, and the reverse direction re-stringifies. Tool results. OpenAI: a separate message with role: "tool" , tool_call_id , and content . Anthropic: a tool_result block inside the next user message's content array, with tool_use_id . Consecutive OpenAI tool messages get coalesced into one Anthropic user message containing multiple tool_result blocks (the shape Anthropic actually wants). Images. OpenAI: {type:'image_url', image_url:{url}} inside a user content array. Anthropic: {type:'image', source:{type:'base64',media_type,data}} or {type:'image', source:{type:'url',url}} . We detect data: URLs and convert them; bare URLs pass through as URL-source images. Plain text content. OpenAI uses a string; Anthropic accepts both a string and a one-element [{type:'text', text}] array. We use whichever is simpler.

OpenAI argument escaping. OpenAI's arguments is a JSON-encoded string, not a parsed object. If it's malformed (rare, but happens with tiny models), we put it under _raw in the Anthropic input instead of throwing. Anthropic top-level system. When converting OpenAI → Anthropic, the result is {system, messages} . If you want just the messages array, use the .messages field — but remember to attach system to your API request separately. Strict alternation. Anthropic requires user / assistant / user / assistant in strict alternation (after the optional system). The converter mostly preserves this, but if your OpenAI source has two consecutive user messages, the Anthropic output will too — and the API will reject it. Merge them first. Multi-modal in OpenAI. An OpenAI user message with images uses a content array instead of a string. We handle this; just make sure your input parses as JSON (no Python-style None /single quotes). Privacy. Nothing leaves the page. Conversion is pure JavaScript; no API keys, no network calls.

OpenAI ⇄ Anthropic Converter

Convert message arrays between OpenAI Chat Completions and Anthropic Messages API formats. Handles system messages, tool calls, image content blocks, multi-turn. In-browser, no API keys.

OpenAI → Anthropic Anthropic → OpenAI

Source (OpenAI Chat Completions)

Target (Anthropic Messages API)

What is this for?

OpenAI's Chat Completions API and Anthropic's Messages API both let you send a multi-turn conversation, but they shape the JSON differently enough that prompts and tool-call traces don't drop in cleanly when you switch providers. The mismatches are small (system message placement, tool-call vs tool-use, image block shape, role: 'tool' vs content-array tool_result) but each one is a 10-minute rabbit hole when you're under deadline. This converter handles the translation so you can paste an OpenAI conversation and get back an Anthropic-shaped one (or vice versa).

What gets translated

System messages. OpenAI puts the system instruction as a regular message with role: "system". Anthropic puts it at the top level as a system field on the request. We move it across and concatenate if you have multiple. (Anthropic accepts a single string for system; multiple OpenAI system messages get joined with double newlines.)
Tool calls (assistant side). OpenAI: tool_calls: [{id, type:'function', function:{name, arguments:'JSON-string'}}]. Anthropic: a tool_use block inside the assistant's content array with {id, name, input} (parsed object, not a string). We parse OpenAI's escaped JSON arguments into the structured input Anthropic expects, and the reverse direction re-stringifies.
Tool results. OpenAI: a separate message with role: "tool", tool_call_id, and content. Anthropic: a tool_result block inside the next user message's content array, with tool_use_id. Consecutive OpenAI tool messages get coalesced into one Anthropic user message containing multiple tool_result blocks (the shape Anthropic actually wants).
Images. OpenAI: {type:'image_url', image_url:{url}} inside a user content array. Anthropic: {type:'image', source:{type:'base64',media_type,data}} or {type:'image', source:{type:'url',url}}. We detect data: URLs and convert them; bare URLs pass through as URL-source images.
Plain text content. OpenAI uses a string; Anthropic accepts both a string and a one-element [{type:'text', text}] array. We use whichever is simpler.

What doesn't translate

Streaming events. This converter handles request/response JSON, not the SSE event streams. If you have a stream dump, reconstruct the full final message first.
Provider-specific fields. Things like OpenAI's logprobs, Anthropic's stop_sequences, sampling parameters, etc. live at the request top level, not inside messages — they're not part of the format we touch.
Cache control hints. Anthropic's cache_control: {type:'ephemeral'} hints have no OpenAI equivalent. They survive going OpenAI → Anthropic (no source to attach them to anyway), but are dropped when going Anthropic → OpenAI.
Tool definitions. The schema of your tools (the tools field on the request) lives outside the messages array. Each provider's tool-definition shape is different (OpenAI: function calling; Anthropic: input_schema). Convert those separately.

Common gotchas

OpenAI argument escaping. OpenAI's arguments is a JSON-encoded string, not a parsed object. If it's malformed (rare, but happens with tiny models), we put it under _raw in the Anthropic input instead of throwing.
Anthropic top-level system. When converting OpenAI → Anthropic, the result is {system, messages}. If you want just the messages array, use the .messages field — but remember to attach system to your API request separately.
Strict alternation. Anthropic requires user / assistant / user / assistant in strict alternation (after the optional system). The converter mostly preserves this, but if your OpenAI source has two consecutive user messages, the Anthropic output will too — and the API will reject it. Merge them first.
Multi-modal in OpenAI. An OpenAI user message with images uses a content array instead of a string. We handle this; just make sure your input parses as JSON (no Python-style None/single quotes).
Privacy. Nothing leaves the page. Conversion is pure JavaScript; no API keys, no network calls.