How to Use Perplexity Research Directly in Codex or Claude Code

You can use Perplexity Research directly in Codex or Claude Code through the Perplexity Research plugin. After setup, ask your assistant to run research, follow up on a finding or export a saved report with its citation mapping. The plugin connects the assistant to your Perplexity session through MCP, the interface assistants use to call tools.

We have added the bridge to the Rhize Plugins marketplace. It is an experimental macOS integration with Perplexity's undocumented web endpoints, not an official Perplexity API integration. Start with a supervised research pilot. It is not a suitable dependency for a critical process that must run unattended.

Why useful research gets lost between tools

Consider a company comparing software vendors. One person collects product pages. Another asks an AI assistant for a comparison. Someone else copies a conclusion into a project brief. By the time the team discusses a recommendation, the original sources and the question that produced it may be difficult to find.

That handoff creates several separate jobs: preserve the question, keep the report, identify which source supports each claim and decide what still needs checking. A polished summary does not do all of them. A missing source can matter more than an awkward paragraph when the decision involves a contract, a customer promise or a new service.

The bridge gives a technical operator a way to retain the research output within an assistant-led workflow. Its potential business benefit is less manual handoff work and a clearer trail back to supporting material. We have verified specific technical behaviors; we have not established a measured company-wide time saving or return on investment.

What the bridge does

The released plugin provides Search and longer-form Research requests, follow-up questions, local job tracking, saved-result retrieval and Markdown export. Jobs retain their state so an operator can inspect a request after the original tool call ends. Once a session is configured, querying does not require operating the browser interface.

The bridge keeps citation mappings explicit. A full report and its shorter summary can have different citation sets. The bridge preserves those mappings separately so the assistant does not treat a summary's source numbers as the full report's source numbers. A human still needs to open important sources and check that they support the conclusion.

It also provides a recovery path for an existing report. If retrieval fails after a request went through, the operator can try to recover that report without submitting the research question again. This matters because a failed local response does not necessarily mean the remote research failed. Automatically asking again could create duplicate work.

The implementation and setup requirements are documented in the public plugin README: https://github.com/Rhize-Media/rhize-plugins/tree/main/perplexity-research

How to set it up and run your first research request

1. Install the Perplexity Research plugin from the Rhize marketplace. In Codex, add the marketplace with codex plugin marketplace add https://github.com/Rhize-Media/rhize-plugins, then run codex plugin add perplexity-research@rhize-plugins. In Claude Code, use /plugin install perplexity-research@rhize-plugins after adding the marketplace.

2. Follow the plugin README to configure Python 3.10+, cryptography and the private JSON settings on macOS. Document your actual authorization and access basis, then restart the host. Ask Codex to set up the plugin, or use /perplexity-research:setup in Claude Code. Connect your signed-in Perplexity session through the supported encrypted handoff and check research_health.

3. Ask: "Use Perplexity Research to compare these three named vendors against these requirements. Include sources and flag unknowns." Requests consume your subscription allowance. The assistant submits a job, checks its status and retrieves the report. Keep the same job when checking progress; a lost response is not a reason to submit again.

4. Ask a specific follow-up, such as "Which integration claims still need vendor confirmation?" Then request a Markdown export. Review the decision-critical sources before using the findings. The setup guide remains the source for detailed configuration and host limitations.

Where companies could use it

The following are hypothetical uses, not customer results or tested industry outcomes. Each starts with a bounded question and ends with a person checking the information that affects the decision.

Compare vendors before a shortlist meeting

A small operations team could ask for a comparison of three scheduling platforms against named requirements: supported integrations, data export, onboarding responsibilities and published support options. The useful deliverable is a source-backed comparison with unknowns called out, rather than a confident ranking based on generic features.

A follow-up could narrow the question to one missing requirement. The team could then export the report into its evaluation packet and ask each vendor to confirm the details. The bridge helps carry that packet forward; it cannot verify a vendor's private contract terms or replace a demonstration using the company's actual workflow.

Prepare a market-entry question

A home-service company considering a nearby service area could use AI market research to gather public information about competing offers, advertised service coverage and local customer concerns. The owner should specify the geography, date range and sources that count as credible evidence.

That research can help frame questions for interviews and a small demand test. It cannot establish local demand simply by counting search results, nor can it tell the company whether it has enough field capacity. Keep the public-source findings separate from internal dispatch, margin and customer data.

Prepare a client or leadership brief

A consultancy could collect background on a client's market before a discovery meeting. An internal technical lead could prepare a comparison of integration options before a planning session. In both cases, a saved report gives the next person something concrete to review rather than an unexplained recommendation.

The operator should state which claims were checked, which remain uncertain and what changed after follow-up questions. Avoid including confidential customer information in research requests unless the company's data rules and the provider relationship permit it. Local credential storage does not mean the research itself happens entirely on the local machine.

A practical first pilot

Choose one recurring research task with a known reviewer. Use a question narrow enough that the reviewer can judge whether the answer is useful. Record the current process before changing it: how long preparation takes, how many claims need correction and whether the next person can locate the relevant sources.

Write the request around a decision. For example: "Compare these named products against our listed requirements using current vendor documentation. Distinguish confirmed capabilities, unavailable information and questions for the sales calls. Include source links and dates." Supply only information the company is comfortable sending to Perplexity.

After the report returns, inspect several decision-critical claims against their original sources. Ask a follow-up for a specific gap rather than treating the first report as complete. Export the checked report and record the human review alongside it. If retrieval fails, inspect the existing job and use recovery before considering another submission.

Measure the whole task, including correction and review. A shorter generation time is not a useful win if the reviewer spends longer repairing unsupported statements. Decide whether to continue the pilot based on the quality of the final handoff, not the length or confidence of the generated report.

What setup and safeguards mean in practice

The current bridge requires macOS, a qualifying Python environment and an entitled Perplexity session. It supports an encrypted local browser-session handoff into macOS Keychain. Configuration belongs in a private JSON file; passwords and session cookies do not belong in that file, in prompts or in source control.

The plugin requires a record of the user's actual authorization and access basis before new submissions. Having a subscription should not be represented as evidence of separate provider permission. A technical operator should review the documented access limitations as part of deployment.

The local worker can be cancelled, but local cancellation is not proof that Perplexity stopped its remote work or returned any quota. Undocumented endpoints can change. Treat authentication, provider availability, response correctness and recoverability as separate checks.

Our release verification included retrieval of an existing report through MCP without creating another query. That test covers the recorded recovery path; other reports and provider changes can still cause failures. The recorded runtime evidence is for Codex; it does not establish equivalent live operation in every supported host. Release evidence: https://github.com/Rhize-Media/rhize-plugins/blob/main/docs/release/perplexity-marketplace-2.111.0.md

Start with one research handoff

If your team already uses an AI coding assistant, the Perplexity Research plugin is available through the Rhize Plugins marketplace. Have your technical operator read the setup guide and choose one supervised research task. Keep the first success criterion simple: can the next person understand the finding, inspect its sources and use it without rebuilding the research from scratch?