It starts with a runnable project and a measurable evaluator
Praxist takes over the experimental loop after a team has already made its project run. You define the objective, metrics, constraints, baseline, and evaluator. Its peers then try competing changes, turn evaluation results into stored evidence, and carry selected findings into later generations. That boundary is sensible: the tool can coordinate research, but it cannot decide whether a scientific result is valid for your domain.
The checkout we measured is substantial for a young Python project: 1,237 files, about 306,050 lines of source, and 23.5 MB before installation. Much of that surface covers lifecycle control, evidence storage, scheduling, provider adapters, packaged skills, examples, and task templates. This is closer to a research operating system than a small agent loop. Buyers should expect to learn its concepts before trusting an unattended run.
Codex and Claude Code are operator interfaces, not included runtimes
The README recommends operating Praxist through Codex and also documents Claude Code. A takeover skill inspects a project, creates or repairs the task harness, checks the evaluator and evidence contract, then launches only after its gates pass. Other skills cover onboarding, task initialization, control, diagnostics, literature work, and terminal plots. Direct CLI commands remain available for status, monitoring, stop, resume, examples, and setup.
Installation was quick in our Debian sandbox relative to the system's breadth. It succeeded in 46 seconds, pulled 131 packages, and occupied 905 MB on disk. The documented complete install includes agent and Codex extras, followed by an interactive setup wizard. A saved Codex login can support the Codex-native route. API-backed profiles instead require provider credentials, and sustained parallel research can create a material provider bill that the repository cannot predict for you.
What happened when we ran it
Our run at commit 7af6a26 installed successfully, then the build failed with exit code 1 after 6 seconds. The supplied result does not include the build error text, so there is no defensible cause to report. The narrow finding is still useful: a fresh unprivileged Debian container with 3 CPUs and 8 GB of RAM did not clear the project's build step.
The tests ran for 900 seconds and hit the time limit. Their tail showed many passing cases, several skipped cases, and repeated F markers by 26% progress. It did not report a final total because the suite never finished, and the excerpt does not identify the failing assertions. Pip-audit found 0 known vulnerabilities. That audit result says nothing about the correctness failures or code that the timed-out suite had not reached.
The Fair Source license rules out casual corporate adoption
GitHub could not identify the custom license with an SPDX label, and the repository has no latest GitHub release response. The included agreement permits internal use and modification below its revenue threshold, restricts standalone redistribution, and requires product-name attribution when generated output is released publicly. Organizations reaching $1 million in annual revenue must notify Sapient and negotiate a commercial license under stated deadlines. Legal review belongs near the start of evaluation.
The license question matters more than the 905 MB environment cost. An academic-institution exemption covers specified nonprofit teaching and research uses, while a controlled commercial spin-off does not inherit that exemption. Optional product-usage collection is a separate consent choice. The privacy notice says an unset or denied choice sends nothing, and setup must not treat acceptance of the license or user agreement as telemetry consent.
Linux 3.11 and 3.12 are qualified; Windows-native is excluded
Release CI continuously tests Linux on CPython 3.11 and 3.12. macOS and other CPython 3.11+ hosts are compatibility targets that must pass praxist doctor and task checks. Windows-native sits outside the runtime contract. The package also leaves CUDA, model-training frameworks, datasets, simulators, task containers, and cluster scheduling to the project or host administrator. That makes the advertised one-line install only the Praxist layer.
Our 900-second timeout reinforces the need to test the exact host and task stack rather than treating successful pip installation as readiness. Praxist expects writable run directories and user configuration, and detached runs continue after the foreground monitor closes. Long experiments need resource limits, provider budgets, stop conditions, storage policy, and a baseline that stays reproducible while agents modify candidate copies.
August issue activity is fast, while runtime edges are still moving
The last push was August 30, 2026. GitHub listed 4 open issues and pull requests, and the tracker showed same-day work on a stop-signal path bug. Another open report from August 28 describes a saved Codex login being lost when Praxist stages a separate Codex home. Both reports concern lifecycle or authentication paths that can block real runs, rather than cosmetic gaps. A related Codex Desktop environment fix had already merged.
PRAXIST had 4,484 stars when fetched, but the public issue tracker only began days earlier and offers little history for judging support under load. Five CI workflow files and a tests directory are good project signals. Our checkout still failed its 6-second build and did not finish testing in 900 seconds. Trial it on a disposable, representative project first, and make a clean build plus completed suite your adoption gate.