What a GitHub search for an open-source Perplexity Computer finds

Search GitHub for "perplexity computer" and about 130 repositories come back, none of them Computer itself. The strongest hits are prompt packs that port Computer's skills into other agent tools. A security course uses Perplexity in its labs. There is an unofficial iMessage bridge for messaging Computer from a phone. There is also a clone whose description claims it is modelled on Computer, last committed to in March 2026. Perplexity's own bill of materials does not surface near the top of its own product's name.

That absence tracks the architecture. Computer's planner, the model routing underneath it and the connector catalogue all live inside Perplexity's hosted product; no checkout contains them. What the company did publish, in October 2026, is the execution layer that those tasks run on.

So two repositories are worth an hour, and they sit at different altitudes: SPACE, which executes agent code, and Kortix, which stores the agent.

The Kortix docs spell out the tree a fresh project commits to its default branch (Read the docs):

kortix.yamlAGENTS.mdagents/<name>.mdskills/<name>/SKILL.mdmemory/tools/<name>.ts

The manifest at the root grants capability. AGENTS.md holds standing instructions shared by every agent on the project, and each agent's own behavior sits in a markdown file beside its grants. None of it is a database dump or a vendor export. A diff tells you what changed, and a new host needs no migration wizard.

SPACE, short for Sandboxed Platform for Agentic Code Execution, is Rust code that boots Firecracker virtual machines and exposes them over HTTP. An application asks for a Linux environment from a template, runs commands, streams stdout back, and later pauses or destroys the machine. The checkout carries a full documentation set. There is a local Kind walkthrough, a no-Kubernetes architecture note and an EKS reference deployment. Warm pools pre-boot sandboxes, and shims let code written against the E2B Python SDK run unchanged (repository, read 19 October 2026).

Its README is unusually clear about the boundary. "Your application decides what to run and what results to keep; SPACE manages the sandbox, its processes, and the connection carrying command output."

Kortix is the open-source AI Operating System, and its checkout answers that first half. An agent is agents/<name>.md (behavior) plus an agents.<name> block in kortix.yaml (governance). The manifest names what an agent may call, and its work reaches main as a change request (Kortix on GitHub).

Put them together and the stack is complete. SPACE decides where code executes. Kortix decides which agent runs, what it is allowed to touch, what it remembers between sessions, and what wakes it up on Monday morning. A platform team with its own agent can point it at SPACE. A company that wants the whole thing as text it controls clones Kortix.

Perplexity opened the basement and kept the house. A SPACE checkout gives you an execution environment and an HTTP API. The planner that turns one sentence into a list of steps is still yours to write or buy. Perplexity said as much in the release note, which calls SPACE "a production-tested foundation for agent runtimes" and never claims Computer itself had moved (the announcement).

The checkout that does carry an agent is Kortix. Prompts, skills, accumulated memory, connector wiring and cron schedules all sit in the tree. Open source. Read it, fork it, audit it.

Getting it running takes three commands on a host that already has Docker:

curl -fsSL https://kortix.com/install | bashkortix self-host init --domain kortix.example.com --admin-email you@example.comkortix self-host start

Setup wants an admin address and credentials for one sandbox vendor, and never a model key, because the instance ships no managed models of its own. Git hooks up later from Settings → Git, and models are attached per project under Customize → Models → Providers (Read the docs). The self-hosting guide takes the install step by step.