Repository navigation
Saturn v0.5.1
Pre-releaseSaturn is a workbench for AI research on your own computer: the app, its research assistant and its local runtime in one download.
Preview build. The macOS app is not signed with an Apple Developer ID (Apple Silicon only), and the Windows installer is signed with Saturn's self-signed certificate, which Windows does not trust, so both systems warn the first time you open them.
Install
macOS (Apple Silicon)
- Download
Saturn-0.5.1-mac-arm64-unsigned.dmg, open it, and drag Saturn into Applications (the.ziphas the same app). - Open Saturn. When macOS says it can't verify the developer, go to System Settings → Privacy & Security and click Open Anyway.
Windows (x64)
- Download
Saturn-0.5.1-win-x64-setup.exeand run it. - If SmartScreen stops it, click More info → Run anyway.
- Optional: confirm the installer is Saturn's. In PowerShell,
(Get-AuthenticodeSignature .\Saturn-0.5.1-win-x64-setup.exe).SignerCertificate.Thumbprintshould printD285DA425077C16A3D115D61F9748AFA574BC786.
Then open Settings → Models and connect a model, add a project folder, and write your first question (the first launch walks you through it). The interface is in English or Traditional Chinese, following the system or chosen in Settings → General.
Updating: from a version before the rewrite in Rust and Tauri (0.3.1 or earlier), install this one by hand; your data in ~/.saturn-app keeps working. Later versions update themselves: the update downloads in the background and installs when you quit Saturn, and it must carry a signature from Saturn's updater key.
Check a download (optional): each file has a .sha256 beside it, and a signed build attestation that proves it was built from this repository's source by its public release workflow:
shasum -a 256 -c Saturn-0.5.1-mac-arm64-unsigned.dmg.sha256
gh attestation verify Saturn-0.5.1-mac-arm64-unsigned.dmg -R Saturn-Technology-Project/Saturn-AppThe signed attestations are attached too (*.sigstore.json); with one, gh attestation verify <file> --bundle Saturn-0.5.1-mac-arm64.sigstore.json -R Saturn-Technology-Project/Saturn-App checks a macOS download offline (-win-x64.sigstore.json for Windows).
Saturn-0.5.1-sbom.cdx.json lists the crates, npm packages and model this release is built from (CycloneDX 1.5, with their hashes), for your scanner or your security team.
saturn-worker-0.5.1-linux-x64 and -linux-arm64 are the worker Saturn installs on a Linux SSH host when you press Install worker in Environment; you do not need to download them. Saturn checks each against the hash built into this release before sending it, and the host needs libssl 3 (Ubuntu 22.04, Debian 12, RHEL 9 or later).
Changes
What changed in each version, and whether it affects you, is in CHANGELOG.md.
What changed
Fixed
- Workspace files can be created and renamed from the file tree. New files open immediately, including empty notebooks. Renaming keeps open tabs and unsaved edits, checks an open file's revision, and refuses to overwrite an existing name. F2 renames a focused file.
- The terminal prompt follows its output. Commands appear immediately and output accumulates from top to bottom, with the next prompt underneath, rather than in a footer at the panel's bottom. Up/Down recalls commands and restores an unfinished line; Ctrl+L or
clearclears the scrollback. - A training run that lists its checkpoints can register its analysis outputs again. The checkpoint entries a program writes to its artifacts manifest, for resuming, made "artifacts from run" refuse the whole manifest. They are now passed over, and the figures and JSON beside them are registered.
- Samples changed after a run was sealed are flagged wherever the run is listed. The warning on the Experiments page never showed, because the list of runs did not check the seal; only the run's own Evidence did. The list, the Workspace's run tab and the Samples page now say so too.
- The skills no longer tell a program to write where the sandbox refuses. The checkpoint, pretraining and task skills said to save into the project folder, which a sandboxed run cannot write; they now say
SATURN_RUN_OUTPUT, and a project folder only when the person has turned the sandbox off. The starter program writes its files toSATURN_RUN_OUTPUTas its own rules say, and the list of run variables is complete. - The documents match the runtime on remote runs: the worker is started without nohup, the host checks every file of the version before running (the old "not certified identical" wording is gone from the docs and from the worker's note), the worker is installed from Environment rather than by hand, a run on the host gets only
SATURN_RESUME_FROM, and a project's skill can also live in.agents/skillsor.claude/skills.
Changed
- The source editor's Tab completes, advances snippet fields or indents, with Shift+Tab to go back or outdent, local document-word suggestions alongside language completions, and entry-point/test snippets for Python. Python, JS/TS and JSX, Rust, JSON, Markdown, Shell, YAML, TOML, TeX and Lean have syntax support; JSON also reports syntax errors. Existing indentation and CRLF endings are preserved. Recent source tabs keep undo, cursor and scroll while the Workspace is open, and a status bar shows position, indentation, line endings and editing shortcuts. Project-wide type checking and definition navigation still need a language server.
- The assistant's answers are drawn as Markdown. Headings, lists, quotes, tables, links, bold and inline code show as such instead of as raw marks; fenced code stays a copyable snippet. A link opens in your browser.
- A run's evidence shows more of what it records. Expected: the version's inputs by role, its models with whether each revision pins the weights (a commit or hash) or is a name such as
mainthat can move, and what it resumes from. Observed: the platform or host and its working folder, start, end, duration and exit code, the host's commit, where the sandbox let the program write and how many folders it hid, what a resumed run took over, and whether the snapshot was checked after the run. Reported: the settings and the restore the program described, in words rather than raw JSON, and how many report lines it wrote. The sealed samples' hash, size and count are at the bottom. - An experiment's seeds can be set in Experiments, beside the sandbox, as a comma separated list; they are passed as
SATURN_RUN_SEEDandSATURN_RUN_SEEDSand frozen into each version. A paused run says since when.
Added
- New models appear without waiting for a Saturn update. Once a provider's API key is saved, Settings lists the models that provider offers that key (Claude, OpenAI, Gemini and DeepSeek), Saturn's names first, and leaves out the ones the key cannot use. Without a key, and for Claude Code and Codex, Saturn's built-in list is shown, and "Other model ID…" takes any model ID by hand. The key is sent only to the provider's own address or one a connected model of it already uses. Prices are still Saturn's own, so a model newer than Saturn is shown unpriced on the Usage page.
- Claude Mythos 5.1 can be chosen for the Claude and Claude Code providers, beside Claude Fable 5.1. It is the same model, offered by Anthropic to verified organizations, and is priced the same on the Usage page.
- Local language servers in the source editor. Python uses Basedpyright or Pyright with the project's selected interpreter; Environment has a Python language server kit. JS/TS use TypeScript 7's native LSP (project first, then PATH), with the TypeScript 6 language-server wrapper supported; Rust uses rust-analyzer. Connected files have semantic completion, type hover, signatures, versioned push/pull diagnostics and F12 definitions within the project. Unsaved imports survive tab switches and connection restarts. The status bar reports connection state and problems, with restart/setup controls. The runtime owns and reaps server process groups, checks document paths and rejects executable commands and workspace edits. Notebook cells and Windows are not covered yet.
- Programs to start from, and an SDK.
research_templatehands the assistantsaturn_run.py, a dependency-free file an experiment imports to write samples, metrics, its report, checkpoints and artifacts the way Saturn reads them (whole lines, flushed, no duplicate sample, seeds applied and reported, SIGTERM handled), and eight runnable templates to copy into the project and change: a benchmark harness wrapper, a red-team harness, an agent trajectory recorder, a Lean proof checker, a linear-probe trainer, a steering-vector sweep, a resumable training loop, and a paper compile script. A copy is the project's own program, frozen with each version. The SDK is tested against the real Python. - The assistant reads less per turn. The research tools' descriptions are a sixth shorter, with the detail in the
saturn-research-toolsskill; a log gives its last 8 KiB and an output file 32 KiB unless more is asked for; a run's evidence gives the snapshot's file count instead of listing thousands of files unless asked; andstatuscuts a finding's text at 240 characters. - The research tools cover what the skills teach. A
documentartifact keeps a PDF, a model card, a paper or a compile log with its hash, previewed in Experiments (a Markdown card drawn, a PDF shown).research_read what: "output"lists and reads what a run's program wrote toSATURN_RUN_OUTPUT, such as an agent's trajectories or a proof checker's messages, without running anything. A label can carry arater, andwhat: "agreement"(and the Samples page) gives Cohen's and Fleiss' kappa between raters, against the adjudicated labels and against the scorer.research_inspectactivations cansaveevery layer's raw residual stream as an.npzin the project, for probes and steering directions to be trained on in a run. The environment check now reports which programs are on the interpreter's PATH (git, uv, elan, lake, lean, tectonic, latexmk, pdflatex, pandoc), shown in Environment, so a Lean or LaTeX run is promised only when its toolchain is there. - Formulas are set in the assistant's answers.
$…$,$$…$$,\(…\)and\[…\]in an answer are drawn as mathematics (by KaTeX, with nothing in a formula trusted); a formula KaTeX cannot read is shown as written. Two prices in a sentence are not a formula. - TeX and Lean files are highlighted in the Workspace.
.tex,.sty,.clsand.bibshow commands, environments, comments and formulas apart;.leanshows declarations, tactics, nested comments,#commands, andsorryunderlined as what it is. Both readers are Saturn's own. - A skill for Lean 4 proofs,
lean-proofs: checking proofs as runs, evaluating a model that writes them, and accepting nothing withsorry, a new axiom or a statement nobody read. The paper-writing skill says how to compile a LaTeX draft as a run, into the run's output, so the PDF is tied to the sources that made it. - Seventeen more bundled skills for the research assistant, on what a safety lab evaluates and how it finishes work: red teaming, safety benchmarks, behavioral evaluations (sycophancy, deception, sandbagging, reward hacking), agentic evaluations; linear probes, steering vectors, circuit discovery and the tuned lens; reward models and RL, ablation studies, scaling and efficiency (scaling fits, distillation, quantization) and model cards; human evaluation with inter-rater agreement, replication review, data provenance and paper writing.
What's Changed
- Research tools for every skill, experiment templates, Markdown and formulas, provider model lists, and Workspace editing by @rickwengdev in #78
- Saturn 0.5.1 by @rickwengdev in #79
Full Changelog: v0.5.0...v0.5.1