Does anyone try to use Perplexity API on Openclaw? Couldn't do it

General
1

I have try many time’s and different ways even with perplexity computer but it doesnt works at all… Mabe anyone do it already and have some setup tips for me

2

Hey! Have you read Perplexity with OpenClaw - Perplexity?

3

Yeah i see about search but i mean use as LLM model but have a problem, gpts does work at all, sonar same, claude works but only chat mode tools doesn’t work

4

[FIX] OpenClaw + Perplexity Agent API: all tool calls fail with empty arguments {}

OpenClaw version: 2026.4.9
Provider: Perplexity Agent API (https://api.perplexity.ai/v1)
API type: openai-responses
Model: perplexity/anthropic/claude-sonnet-4-6

Symptom

When using OpenClaw with the Perplexity Agent API in openai-responses mode, any tool (exec, write, read, etc.) fails with a validation error:

text

**Validation failed for tool "exec":
- command: must have required property 'command'

Received arguments:
{}**

Chat replies without tool calls work fine — only tool calling is broken. Arguments always come through as empty {}.

Cause

The streaming parser for the OpenAI Responses API (openai-responses-shared.js from @mariozechner/pi-ai) expects function call arguments to arrive incrementally via response.function_call_arguments.delta events.
However, Perplexity Agent API sends all arguments immediately in the first response.output_item.added event (inside item.arguments) and never emits subsequent delta or done events for function arguments.

Here’s what happens step by step:

  1. On receiving response.output_item.added with item.type === "function_call", the parser creates a currentBlock with arguments: {} and stores item.arguments in partialJson.

  2. It then waits for response.function_call_arguments.delta events to fill arguments from partialJson — but Perplexity never sends those.

  3. When response.output_item.done comes, the parser creates a new toolCall object with correctly parsed arguments from partialJson and passes it in the stream event toolcall_end, but does not update the existing entry in output.content.

  4. The agent loop (agent-loop.js in @mariozechner/pi-agent-core) reads tool calls from message.content — where arguments are still {}.

  5. Validation fails since required properties are missing.

Example from the session file:

json

{
"type": "toolCall",
"name": "exec",
"arguments": {},
"partialJson": "{\\"command\\": \\"echo \\\\\\"tools test ok\\\\\\"\\"}"
}

The arguments were correctly captured in partialJson, but never copied to arguments.

Fix

File to patch

text

/usr/lib/node_modules/openclaw/node_modules/@mariozechner/pi-ai/dist/providers/openai-responses-shared.js

What to change

Find the response.output_item.done handler for type function_call (around lines 392–405). It looks like this:

javascript

else if (item.type === "function_call") {
const args = currentBlock?.type === "toolCall" && currentBlock.partialJson
? parseStreamingJson(currentBlock.partialJson)
: parseStreamingJson(item.arguments || "{}");
const toolCall = {
type: "toolCall",
id: `${item.call_id}|${item.id}`,
name: item.name,
arguments: args,
};
currentBlock = null;
stream.push({ type: "toolcall_end", contentIndex: blockIndex(), toolCall, partial: output });
}

Add two lines before currentBlock = null;:

javascript

else if (item.type === "function_call") {
const args = currentBlock?.type === "toolCall" && currentBlock.partialJson
? parseStreamingJson(currentBlock.partialJson)
: parseStreamingJson(item.arguments || "{}");
const toolCall = {
type: "toolCall",
id: `${item.call_id}|${item.id}`,
name: item.name,
arguments: args,
};
// FIX: update entry in output.content so agent-loop sees correct arguments**
const idx = output.content.indexOf(currentBlock);
if (idx >= 0) { output.content[idx] = toolCall; }
currentBlock = null;
stream.push({ type: "toolcall_end", contentIndex: blockIndex(), toolCall, partial: output });
}**

Restart gateway

After saving the file, restart the gateway service:

bash

systemctl --user restart openclaw-gateway

Why it works

The toolcall_end event already contains the correct toolCall object with parsed arguments.
The problem is that the agent loop reads tool calls from message.content, which still holds the old currentBlock with empty {} arguments.
The patch replaces that outdated record with the properly parsed toolCall before clearing currentBlock.

Warning about updates

This patch modifies a file inside node_modules.
Whenever OpenClaw is updated (npm update -g openclaw), the patch will be overwritten — you’ll need to reapply it until it’s merged upstream in @mariozechner/pi-ai.

Extra: proper configuration for Perplexity Agent API

Configuration file: ~/.openclaw/openclaw.json

Parameter Correct value Common mistake
apiType openai-responses openai-completions — doesn’t work
baseUrl https://api.perplexity.ai/v1 /v1/agent — invalid path
model perplexity/anthropic/claude-sonnet-4-6 —

Note: baseUrl must be https://api.perplexity.ai/v1, not /v1/agent.
The API type should be openai-responses.