Open source Perplexity Computer: own the agent in your repo

No. Perplexity has not published an open-source Perplexity Computer. What exists is a field of thirteen open-source projects that do the work, including Kortix, OpenClaw, Eigent, Open WebUI, AnythingLLM, Vane, Khoj and Morphic. The question of whether Perplexity's own product is open source has its own page.

Perplexity published SPACE, the sandbox platform underneath Computer, on 15 October 2026 (SPACE). SPACE runs agent and user code in isolated Firecracker microVMs and is Apache-2.0. It tested its own boundary and reports no escape to the host in 108 published attempts (the announcement). SPACE answers how a task is isolated. What it leaves open is what the task is, which tool it should use and what happens to the result.

The planner, the per-query model routing, the 400+ connectors and the product around them stayed inside Perplexity's product (the announcement). That is the layer a team has to own if it wants to change how the agent works, and it is the layer Kortix puts in a git repo. A team that cannot read the planner can ask the vendor for a change; a team that holds the file can edit it and review the diff. The comparison of all thirteen projects is on the alternatives page.

A GitHub search for "perplexity computer" returned 160 repositories on 2 November 2026 (the query). The narrower search "perplexity space" returned 54 repositories the same day (the query). Perplexity's own account, perplexityai, publishes 57 public repositories (the account).

SPACE is Apache-2.0 in Rust and was last pushed on 2 November 2026 (SPACE). pplx-garden is MIT in Rust, calls itself "Perplexity open source garden for inference technology", and was last pushed on 2 September 2026 (pplx-garden).

The search returns a name in use across many projects, and the count moves. Perplexity's own account publishes the sandbox and pplx-garden, an inference garden, and the census page shows what that leaves out: the planner, the routing and the connectors; Portable Computer is one local route, on a Windows or Linux GPU machine you buy, and Personal Computer's local model on an Apple silicon Mac is the other.

The rival rows come from each project's own repository or site, the first seven read on 3 November 2026 and the five new rows on 7 November 2026: OpenClaw, Eigent, Open WebUI, AnythingLLM, Vane, Khoj, Morphic, GPT Researcher, Local Deep Research, Scira, OpenHands and Open Interpreter.

Kortix is my pick because the agent, its skills, memory and connectors sit in one repo the team owns, and a change to how it behaves arrives as a diff somebody reads; I work on Kortix's marketing, so weigh that accordingly. The boundary that matters is where the agent's definition lives: Kortix keeps it in git, OpenClaw in a gateway config, Eigent and AnythingLLM in an app, Open WebUI in a web admin, Vane, Khoj and Morphic in a Docker or environment file, GPT Researcher and Scira in an environment file, Local Deep Research in a web UI, OpenHands in launch variables and a UI, and Open Interpreter under ~/.openinterpreter. When the definition is a file, a prompt change and the grant it needs can arrive in one diff, and the reviewer sees both before either takes effect.

The starter kortix.yaml is the manifest a new project begins with, and its head is short enough to read in one screen. This is the file, read on 7 November 2026.

# yaml-language-server: $schema=https://kortix.com/schema/kortix.schema.json## {{projectName}}: the Kortix workspace manifest. Agents, what each may use,# triggers and sandboxes live here, committed beside your files.# Reference: https://kortix.com/docs/developers#kortixyaml-referenceproject:  name: "{{projectName}}"  description: A Kortix workspace.# Built-in agents. My Kortix, the agent you talk to, routes each task to the# right agent below. Delete this key or set it to false to leave it out.builtin_agents:  my_kortix: true# The agent a session or trigger runs when it names none. My Kortix starts# `worker` when it names no agent.default_agent: mykortix

Kortix's manifest names the built-in agent in its own block, and a session or trigger that names no agent runs mykortix, which routes the task to the agent that fits it. The agent's prompt, skills, memory and connector grants sit in the same repository, so a reviewer can read the change before it takes effect. The contrast is the layer Perplexity kept closed: Kortix's agent definition and its grants are files your team can open, review and change.

Kortix runs as one Docker Compose stack on a Linux server you control (Kortix docs). After the CLI is installed, the three commands are:

kortix self-host initkortix self-host configurekortix self-host start

Setup asks for reachability, an admin email, a sandbox provider key and the update policy, with no model key among them; GitHub connects later from the dashboard, and models are BYOK in the app (the self-hosting guide).