Perplexity Agent

perplexity-agent-mcp

[CI]

A single-file, zero-third-party-dependency MCP server for Perplexity's Agent API — multi-step web research with citations, exposed to any MCP client over stdio.

Python standard library only: no MCP SDK, no HTTP library, no code generated at build time. perplexity_agent_mcp.py holds your API key and talks to the network on your behalf, so it is written to be read — 1,774 lines, comments included. Both install paths below ship the exact same file; which one you pick changes how many other parties you're trusting to get it onto your disk, not what actually runs.

Background reading: perplexity-agent-mcp: an open-source MCP server for Perplexity's Agent API — why this exists, what the Agent API does that Sonar doesn't, and what building it turned up.

Why this exists

Perplexity's own @perplexity-ai/mcp-server wraps the Sonar chat models and the Search API. It does not reach the Agent API (POST /v1/agent) — Perplexity's multi-step, tool-using research endpoint, which runs its own web searches, fetches pages, and synthesizes one cited answer. This server fills exactly that gap. It's meant to run alongside the official one, not replace it.

The three tools

The Agent API is asynchronous under the hood (see Configuration for why that matters), so this server exposes it as one lifecycle, not one call:

One deliberate asymmetry, visible in tools/list: perplexity_agent's readOnlyHint annotation is false, even though it never touches your local machine — it creates real, billable, cancellable state on Perplexity's servers, and MCP clients use that hint to decide whether a call needs your approval before running. perplexity_agent_cancel carries destructiveHint: true, honestly: it ends a run, and Perplexity reports no usage at all for cancelled runs, so this tool cannot tell you whether cancelling changed your bill (see Security).

Install — Path A: single file

The whole point of this project is that you can read the one file that will hold your API key before you trust it with one. Do that first: download perplexity_agent_mcp.py and read it top to bottom.

Then check your interpreter. Stock macOS ships Python 3.9.6 at /usr/bin/python3 (confirmed on a current macOS install) — below this server's 3.10 floor. It will refuse to start and say why rather than fail obscurely, but you still want a real 3.10+ python3 (Homebrew, python.org, or uv python install all provide one):

python3 --version   # must be >= 3.10which python3       # note this path for the config below

Point your MCP client at the absolute paths of both the interpreter and the file:

{  "mcpServers": {    "perplexity-agent": {      "command": "/absolute/path/to/python3",      "args": ["/absolute/path/to/perplexity_agent_mcp.py"],      "env": { "PERPLEXITY_API_KEY": "pplx-…" }    }  }}

Use absolute paths for both — the same reason Path B insists on one below applies here too: a GUI app's PATH is not your shell's PATH, and a bare "command": "python3" can quietly resolve to that too-old system interpreter instead of the one you meant. On macOS, Claude Desktop's config file lives at ~/Library/Application Support/Claude/claude_desktop_config.json.

Trust chain: Python standard library, plus Perplexity. Nothing else.

Install — Path B: uvx from PyPI

One JSON snippet, nothing to download by hand. uv fetches and runs the published package each time your MCP client starts the server.

{  "mcpServers": {    "perplexity-agent": {      "command": "/absolute/path/to/uvx",      "args": ["perplexity-agent-mcp"],      "env": { "PERPLEXITY_API_KEY": "pplx-…" }    }  }}

Two things worth getting right, each the difference between a working config and a support thread:

Use an absolute path to uvx. macOS GUI apps — Claude Desktop launched from Finder or Spotlight, not a terminal — do not inherit your shell's PATH. If the config above says "command": "uvx", Claude Desktop very likely can't find it and fails with spawn uvx ENOENT. Run which uvx in your terminal and paste the absolute path it prints into command instead.

Pin the version if you want reproducibility. "args": ["[email protected]"] holds you on one release. Unpinned, uv resolves the newest published release on every restart — which is a materially different risk from tracking a branch: releases are immutable, tagged, and go through the same CI as everything else. Pin if you would rather review each upgrade; leave it off if you would rather get fixes automatically. Either is defensible, which is why this is not a warning.

Also: a plugin for llm

The same Perplexity Agent client, exposed as a model for Simon Willison's llm CLI. Optional — the MCP server never imports it and keeps its zero dependencies either way.

llm install llm-perplexity-agentllm keys set perplexity        # skip if you already set this for llm-perplexityllm -m perplexity-agent 'What changed in MCP 2026-07-28?'

Options mirror the MCP tool:

llm -m perplexity-agent -o preset xhigh -o recency week 'Latest on X'llm -m perplexity-agent -o domains 'nasa.gov,-reddit.com' 'Artemis status'llm -m perplexity-agent -o timeout 600 'Something genuinely deep'

Poll progress goes to stderr, so the answer pipes cleanly:

$ llm -m perplexity-agent 'What is MCP?' > answer.md[perplexity-agent] status queued after 1s; no intermediate results yet[perplexity-agent] status in_progress after 3s; 10 search result(s) gathered

Why not just point llm at the MCP server? llm has no MCP support — simonw/llm#696 has been open since January 2025. Third-party bridges exist, but the maintained ones pull in the full MCP SDK to talk to a server that deliberately has no dependencies. And the existing llm-perplexity plugin wraps the older Sonar chat models, not the Agent API — the same gap this project fills for MCP.

One deliberate difference from the MCP server: spotlighting is off by default here. In MCP the answer goes straight into a model that is holding tools, so injected instructions could cause actions, and the wrapper is essential. On the command line the answer goes to a terminal for a human to read, and llm runs no tool loop by default — so the realistic risk is a manipulated summary if you pipe it into another model, not a hijacked agent. Turn it on with -o spotlight true when the output is headed somewhere that matters:

llm -m perplexity-agent -o spotlight true 'Research X' | llm -m gpt-5 'Summarise'

Trust chain: Path A vs Path B

These are not two equally convenient ways to install the same thing. Path B has a strictly larger trust surface than Path A — full stop; that's not a knock on Path B, it's the trade you make for not downloading anything by hand.

Both paths ship the same bytes — flit_core copies the one .py file into a wheel; nothing is generated, bundled, or vendored in between. Reading the file audits both.

Configuration

Why 55

Claude Desktop enforces a tool-call timeout of 60 seconds that its users cannot configure — no setting, no environment variable, nothing in the UI. A synchronous MCP call that runs longer simply dies from the client's point of view, while Perplexity keeps working — and billing — server-side regardless. 55 leaves this server's own submit-then-poll round trip a 5-second margin inside that hard ceiling.

If your client isn't Claude Desktop, raise it — there's no reason to hand back a response_id early if your client is happy to keep waiting:

"env": { "PERPLEXITY_API_KEY": "pplx-…", "PERPLEXITY_AGENT_WAIT_SECONDS": "300" }

300 is what we recommend for Claude Code, VS Code, and Cursor IDE — see why in the table below. An unset, empty, non-numeric, zero, or negative value silently falls back to 55; a valid positive value is capped at 1800 (30 minutes) so a stray extra zero can't hang a call indefinitely. perplexity_agent_result's own wait_seconds parameter advertises this same ceiling as its declared maximum in tools/list — but sending it a larger value anyway doesn't get rejected, it gets silently clamped to the ceiling.

No MCP client publishes Claude Desktop's figure anywhere; we obtained it by reverse-engineering the installed app (v1.24012.1) — its bundled TypeScript SDK defaults DEFAULT_REQUEST_TIMEOUT_MSEC to 60000ms, and Claude Desktop calls callTool() without overriding it. The rest of this table comes from each project's own docs or source, with confidence noted where it's shaky:

Progress notifications do not rescue Claude Desktop: the spec lets a client reset its timeout clock when progress arrives, but only if the client supplied a progressToken in the first place, and Claude Desktop never does. This server's notifications/progress support is opportunistic — pure upside for the other clients, a no-op here.

Protocol revisions

This server speaks both eras of MCP, and works out which one you want from what you send. There is nothing to configure, and no way to get it wrong.

2026-07-28 went GA on 2026-07-28 and removes the initialize/notifications/initialized handshake entirely: the protocol becomes stateless, every request carries its own protocol version and client capabilities, and servers must implement server/discover. It is the largest breaking change in MCP's history.

Nothing about your setup needs to change. A legacy client sees byte-identical behaviour to before — that is asserted by exact-equality tests, not by intention. As of this writing no MCP client has publicly shipped 2026-07-28 support, so in practice every client still takes the legacy path; the modern one is here so that none of them has to wait for this server when they do.

Two behaviours worth knowing if you are testing against it:

• server/discover is answered even without _meta. On stdio it is the designated backward-compatibility probe, and a client reads the reply to decide which era this server is. Any error there would tell it "legacy server" permanently.
• A modern request naming a revision we don't speak gets -32022, carrying the list it should retry with. That is deliberately the opposite of initialize, which never errors on version negotiation and downgrades instead — the two rules govern two different code paths and are not in conflict.

Presets

preset selects research depth. It's an open string on Perplexity's side too — their API reference types it as a plain string, not an enum — which is why this server doesn't constrain it either: a preset Perplexity adds after this README is written still works, unmodified. These are the ones documented today, roughly shallowest to deepest:

The one hard number we have: a single medium run — a three-source comparison query — measured 12.5 seconds and $0.039. That's one sample, not a benchmark; different questions and deeper presets will vary, likely a lot. This server always submits in Perplexity's background mode regardless of preset — poll, don't hold a socket open for the length of a research run — which is also the mode wide-research requires.

See Perplexity's presets guide for the current, authoritative list.

Self-test

No API key needed — none of these calls touches the network. There are two of them because this server speaks two protocol revisions, and they exercise one each.

Legacy (2025-11-25), the handshake era:

printf '%s\n' \  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-11-25","capabilities":{},"clientInfo":{"name":"manual","version":"1"}}}' \  '{"jsonrpc":"2.0","method":"notifications/initialized"}' \  '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' \  | python3 perplexity_agent_mcp.py

Expect exactly two JSON lines back: an initialize result, then all three tools from tools/list. Nothing for the notifications/initialized line — the spec forbids replying to notifications, including ones the server doesn't specifically act on.

Modern (2026-07-28), the stateless era:

printf '%s\n' \  '{"jsonrpc":"2.0","id":1,"method":"server/discover"}' \  '{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{"_meta":{"io.modelcontextprotocol/protocolVersion":"2026-07-28","io.modelcontextprotocol/clientCapabilities":{}}}}' \  | python3 perplexity_agent_mcp.py

Expect two lines again: a DiscoverResult listing every revision this server speaks, newest first, then the same three tools — this time carrying resultType, ttlMs, cacheScope, and a _meta naming the server. Note that the first line has no _meta at all: server/discover is answered unconditionally, because on stdio it doubles as the probe a dual-era client uses to find out which era it is talking to.

Security

Three things worth stating plainly rather than leaving implied.

1. Perplexity retains your queries. This server never sends a store parameter, so Perplexity's own default applies. Perplexity documents store: false as hiding a response from later retrieve calls — and this server's entire design is submit-then-poll-later, for every request, regardless of preset. So storage can't meaningfully be off for how this tool works: assume every query and answer is retained on Perplexity's servers, governed by their retention policy, not this project's.

2. Returned web content is a prompt-injection vector. Every tool here feeds text scraped from the live web to a language model. That's inherent to the category, not specific to this server, and most web-search MCP servers don't say so. Mitigation: answers are wrapped in a randomized delimiter — <untrusted-web-content-{16 hex chars}>…</untrusted-web-content-{same}>, 64 bits of entropy from secrets.token_hex(8) — with an explicit "this is data, not instructions" instruction, and any occurrence of the closing tag inside the content itself is stripped first so a hostile page can't forge its own way out. This is spotlighting by delimiting (Microsoft Research, arXiv:2403.14720). It is a mitigation, not a fix: no client is obliged to honour the delimiter, and no model is guaranteed to respect it.

3. Your API key sits in plaintext in your MCP client's config file. The MCP specification directs stdio-transport servers to skip its OAuth flow entirely and get credentials from the environment instead, and every MCP client implements that as a plaintext env block in a JSON config file on disk. That's the ecosystem's normal, spec-sanctioned posture for local servers, not a shortcut this project took — but it deserves saying rather than assuming you already know it. Treat that config file like any other secret: correct file permissions, never commit it, never paste it into a support ticket.

This server has no other secret-handling surface worth auditing beyond those three: the key is read from the environment at call time, not import time, so a missing key is a clean tool error rather than a crash; it lives only inside one function's local scope; it's never logged; and there is deliberately no way to redirect the API host via an environment variable — that would itself be an exfiltration path in a tool whose entire configuration is a user-editable JSON file.

Development

No install step — that's the point:

git clone https://github.com/zalez/perplexity-agent-mcp.gitcd perplexity-agent-mcppython3 -m unittest discover

190 tests, pure standard library — tests/fake_perplexity.py runs a fake Perplexity over http.server, in-process, so nothing real is ever called.

Before committing, install the git hooks once and let them run automatically after that:

uv tool install pre-commit   # or: pip install pre-commit / pipx install pre-commitpre-commit installpre-commit run --all-files

18 hooks: file hygiene, ruff + mypy --strict, gitleaks and zizmor for secret and GitHub Actions security scanning, codespell, and the full test suite. CI runs the same hook set (see ci.yml) plus a matrix across Python 3.10–3.14 and a packaging job that builds a wheel, installs it into a clean environment, and drives the console script over real pipes — a successful build alone doesn't prove the entry point works.

More

• Design spec — every decision with its rationale, the places the original brief turned out to be wrong, and the live verification results.
• Dual-era MCP — why this server speaks two protocol revisions at once, and the three precedence rules that decide which one your request gets.
• Blog post — the story behind the project.
• SECURITY.md — threat model, and an honest account of what the prompt-injection mitigation does and does not do.

License

BSD-3-Clause. Copyright (c) 2026, Constantin Gonzalez. See LICENSE.