Model or dataset
sapientinc/PRAXIST avatar
sapientinc/PRAXIST

Praxist: an autonomous research loop wrapped around a project you already run

Autonomous research system for measurable, computer-executable research.

5,532 stars733 forksPythonNOASSERTION

At a glance

What is it?
Praxist is a Python system that runs parallel research peers, evaluation, and evidence handling over a project you already have, driven through Codex, Claude Code, or its own CLI. It ships as version 0.5.0, ships no GitHub releases, and refuses to invent the part you were supposed to supply.
Who is it for?
Praxist fits a project that already runs, has an objective you can measure, and can host a long unattended loop. It does not fit anyone looking for a research assistant to define the question, write the evaluator, or pick the metric, because those belong to the task project and the takeover gates will not launch without them.
Can I use it commercially?
Check first. The repository uses a licence we do not classify automatically, so read its LICENSE file before any commercial use.
Is it still maintained?
Yes. The repository last received commits 53 days ago.
What is it written in?
Mainly Python, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • sapientinc/PRAXIST5,532 ★ · -965
Star history
Repository2026-09-212026-11-15
sapientinc/PRAXIST6.5k5.5k

-965 Stars in the last 55 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

One pip command with two extras, and a wizard that stops before the run

Praxist installs from PyPI with a single command that pulls the agents and codex extras and then hands you to an interactive setup wizard:

bash
python3 -m pip install --index-url https://pypi.org/simple "praxist[agents,codex]" && praxist setup --interactive --install-skills codex

That one line covers the complete runtime integrations and first-use setup. The wizard itself handles the Fair Source License, the User Agreement, privacy, the runtime profile, masked credentials, Codex skills, writable examples, and readiness checks. What it will not do is select a research project or launch a run, so a first session ends with an installed tool and a set of checks passed, not with research moving. Two documents are named as the things to read before starting research: the Quickstart and Your First Task, which cover the separate takeover step and the project contract it creates. Claude Code users get a different route, a host-specific one-line command kept in the installation page of the docs rather than in this repository's root.

The second install path hands setup to a session running with yolo

Agent-managed installation is the alternative route. Open Codex with:

bash
codex --yolo

then ask it to install and configure Praxist using the packaged OOBE runbook and to stop after readiness checks. Two properties of that path deserve naming before you use it. The session runs with yolo, so nothing inside it intercepts writes to your machine, and the only stopping condition described anywhere in that flow is the readiness check. The second is authentication, and you pick one: a saved Codex login for Codex-native mode, which uses an existing Codex subscription and needs no API key, or a supported provider API key. On that choice the project says Praxist generally favors open-source model APIs for sustained research, on the strength of a high observed cache-hit rate, while Codex-native mode exists for the case where no key is available. The setup wizard also supports other API-backed profiles.

Four lifecycle commands, and Ctrl-C closes only the monitor

The CLI surface for running work is short:

bash
praxist status --json
praxist --monitor --latest
praxist stop <run_id>
praxist resume <run_dir>

Two of those lines carry traps. Monitoring is a separate view, and Ctrl-C closes only the monitor, not the research run, so stopping work means running the stop command with its run id rather than reaching for the keyboard. Resuming takes a run directory instead of an id, so the path on disk is the handle you keep. To get something runnable to point the loop at, the same repository ships example projects:

bash
praxist examples list
praxist examples install rocket_booster_recovery
praxist examples install rocket_booster_recovery_rust

Complete examples are writable reference projects rather than read-only demos, and `templates/tasks/` holds replaceable scaffolding for building new task harnesses.

Nine bundled skills, and the takeover brief decides whether a run starts

Skills are the interface layer; besides praxist-takeover, the README lists these other bundled skills:

| Skill | Purpose | |---|---| | `praxist-takeover-codex` | No-key takeover using the saved Codex login | | `praxist-onboarding` | Explain Praxist and inspect local readiness | | `praxist-task-initialization` | Build or repair a task harness without launching | | `praxist-interactive-task-init` | Design a task through confirmation-first setup | | `praxist-control` | Start, stop, resume, monitor, and inspect runs | | `praxist-diagnostic` | Diagnose run health and produce reports | | `praxist-scientific-research` | Gather sourced literature and benchmark context | | `praxist-runtime-install` | Install or repair runtime dependencies and credentials | | `terminal-line-plot` | Draw metric trends in the terminal |

The takeover skill is the entry point. After installation you open Codex in the root of an already runnable research project and invoke `$praxist-takeover`; it inspects readiness, creates or repairs the task harness, validates its evaluator and evidence contract, and launches the run once the required gates pass. What gets launched comes from the brief you write. The example brief asks for the objective with a metric and direction, the constraints to preserve, a peer count, a generation count, a time or cost budget, the runtime and model provider chosen at setup, whether to allow or disable literature search, whether to enable QD, and whether generation-zero DIG is on. The project's own note is that a precise brief produces a better research plan, and no gate fills in missing numbers for you.

Praxist owns the loop while your project stays the source of truth

The ownership split is worth reading before anything else:

| Praxist owns | The task project owns | |---|---| | Research orchestration, lifecycle, evidence protocols, replay, scheduling, and extension interfaces | Research objective, executable code, evaluator, metrics, baselines, prompts, roles, and domain constraints |

The line underneath it decides whether the tool fits: Praxist contains no task-specific scientific assumptions, and a task remains the single source of truth for what should be tested and what counts as valid evidence. The consequence for a reader is direct. If the project does not already run, or the objective is not measurable, the orchestration half has nothing to orchestrate. The requirement list is blunt about it, listing a runnable project with measurable evaluation as required to launch research, and the takeover skill validates the evaluator and the evidence contract before anything starts. Behind that boundary sit parallel research peers for competing hypotheses, multi-generation synthesis, durable evidence lanes through incubator, frontier, and Gems state, multi-metric ranking that includes Pareto-optimal tradeoffs, Quality-Diversity with an optional Deep Innovation Gate, central resource scheduling, and resume, replay, and monitoring.

The wheel declares three dependencies, requirements.txt declares two dozen

What pip installs is narrower than what the repository works with. The packaged dependency list is three names:

toml
dependencies = [
    "pyyaml>=6.0",
    "jinja2>=3.1",
    "pydantic>=2.7,<3",
]

Everything else is optional. The agents extra pulls claude-agent-sdk at a pinned 0.2.136, mcp, anthropic, openai, httpx, requests, and a document extraction stack of trafilatura, pdf2image, pytesseract, and pypdf. The codex extra pins openai-codex at 0.147.0 and codex-relay at 0.5.5. Storage is boto3 plus botocore. A product-usage-server extra brings alembic, fastapi, psycopg with the binary extra, sqlalchemy, and uvicorn, while docs brings mkdocs-material and mkdocstrings. Set against that, requirements.txt is unpinned and far wider: anthropic, openai, transformers, datasets, wandb, numpy, matplotlib, seaborn, tiktoken, black, python-igraph, omegaconf, and jsonschema among them. The gap tells you the install command hands you orchestration rather than a training stack, and that file describes a working environment instead of the package contract.

Four console scripts, and release testing stops at Linux

pyproject.toml registers four entry points:

toml
[project.scripts]
praxist = "praxist.cli:main"
praxist-uninstall = "praxist.cli.uninstall:main"
praxist-collector = "praxist.product_usage.app:main"
praxist-retention = "praxist.product_usage.retention:main"

The collector and retention scripts sit on a product_usage module, which is what the service extra with alembic, fastapi, and psycopg exists for, and the tree carries a services/ directory beside scripts/, skills/, templates/, tests/, and docs/. There is a shell side too, with praxist-install.sh and praxist-uninstall.sh next to pyproject.toml, plus AGENTS.md, mkdocs.yml, a .python-version file, and a .coveragerc. Two facts about how this project ships shape your evaluation. The version in pyproject.toml is 0.5.0 while the repository carries no GitHub releases, so the release list will not tell you what version reached PyPI. And the support matrix has an honest width to it: CPython 3.11+ is required, release testing runs on Linux with CPython 3.11 and 3.12, and macOS plus other 3.11+ environments are a compatibility target where the docs send you to `praxist doctor` first. License terms live in LICENSE.md, referenced as a file rather than as a named license identifier.

Editorial conclusion

Praxist fits a project that already runs, has an objective you can measure, and can host a long unattended loop. It does not fit anyone looking for a research assistant to define the question, write the evaluator, or pick the metric, because those belong to the task project and the takeover gates will not launch without them. Before installing, confirm the evaluator exists, choose between a saved Codex login and a provider API key, and check your platform against the release-tested row rather than the compatibility row.

Frequently asked questions

What does the Praxist setup wizard do during first use?

It covers the Fair Source License, the User Agreement, privacy, the runtime profile, masked credentials, Codex skills, writable examples, and readiness checks. It does not select a research project and it does not launch a run, so first use ends before any research starts.

Can I run Praxist without Codex or Claude Code?

Yes for direct CLI work. The requirement table marks Codex or Claude Code as needed only for skill-driven operation, and states that direct CLI operation remains available without either. Running the CLI itself still needs CPython 3.11+ and a runnable project with measurable evaluation.

How do I stop or resume a Praxist run?

Stopping takes a run id through `praxist stop <run_id>`, and resuming takes a directory through `praxist resume <run_dir>`. `praxist status --json` and `praxist --monitor --latest` report on runs, and Ctrl-C closes only the monitor rather than the research run.

What do the two Rocket Booster Recovery examples in Praxist show?

They demonstrate the same research problem through a Python/JAX implementation and a native Rust one, and they ship as writable reference projects rather than read-only demos. New task harnesses can start from the replaceable scaffolding in `templates/tasks/`.

Official sources

  1. Issues
  2. Project website
  3. README
  4. sapientinc/PRAXIST on GitHub
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/sapientinc-praxist.svg)](https://hysenlabs.com/projects/sapientinc-praxist)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/sapientinc-praxist.svg)](https://hysenlabs.com/projects/sapientinc-praxist)

Was this useful?