v0.6.0

GraphForge is an embedded, openCypher-compatible graph and analytics engine. It runs inside your Python or Node process, or through the gf command, with no database server. It is pre-v1.

What's in v0.6.0

• Local graphs. Build, load and query graphs locally, with analyst verbs and durable save/reopen. Tabular results are Arrow; graph data is Parquet.
• Research history. Branches, Versions (commits with parents and authors), tags, and git-style CLI verbs: gf branch, gf commit, gf log, gf status, gf push, gf pull.
• GraphForge Hub at graphforge.sh. Browse public Projects, gf clone them with verification, continue locally, and gf publish new Versions back.
• Ontologies. Permanent ont_… module IDs, adopt/fork with provenance, per-module SPDX licenses, and machine-readable descriptors. Start a Project from an exact published ontology with gf project new --from-ontology.
• Portable packages. One verified, native portable format records its producing Core version, with streaming OCI transfer.
• Agent support. Project-local agent skills and the @curatelabs/graphforge-agent-skills package.

Rust owns all behavior. Python and Node are thin bindings.

Install

pip install graphforge            # Python (includes `gf`)
npm install @curatelabs/graphforge      # Node
npm install -g @curatelabs/graphforge-cli   # `gf` CLI
cargo install graphforge-cli      # from source (see Known issues)

Native gf archives for Linux (x86_64, aarch64), macOS (x86_64, arm64) and Windows (x86_64) are attached below, with SHA-256 files. The protocol WebAssembly module is @curatelabs/graphforge-protocol-wasm.

Known issues and limits

• Pre-v1. No backward-compatibility or migration guarantee. Only current-format projects open; others return GF_UNSUPPORTED_PROJECT_FORMAT. GraphForge warns when it opens, imports or clones data from a different Core version.
• crates.io builds (cargo install graphforge-cli) compile against stock DataFusion 54. They lack two fixes that the Python wheel, the npm packages and the GitHub Release archives carry: a spill-pool hang (apache/datafusion#24883) and a join-planning error (#2084). Prefer the release archives, PyPI or npm for heavy queries.
• Hub publishing size. Every graphforge.sh account holds up to 128 MiB of published and pending objects, at most 64 MiB per object, on every plan. Larger Projects are refused with storage_cap; request more storage from your account page. The large-object upload path is post-launch Hub work (CurateLabs/graphforge-nextjs#48).
• Headless publishing. gf auth login and gf publish need an OS credential vault (Keychain, Windows Credential Manager or a Secret Service keyring). Without one they fail closed with hub.auth_denied: OS credential vault unavailable, even in a terminal. In CI and agents, set a repository-scoped GRAPHFORGE_HUB_PUBLISH_TOKEN; on a headless Linux host, run an unlocked Secret Service keyring.
• Reading a Branch head. gf query reads the Project's workspace generation. gf commit records Branch Versions. Read a Branch head with gf research branch query --branch-uuid … or GraphForge.query_research_branch().
• Repository module fetch. gf ontology module fetch OWNER/REPO returns modules in the repository's current head composition. Earlier module versions stay listed, with their licenses, on the ontology's Hub pages.
• Unproven filesystem classes open with GF_UNVERIFIED_FILESYSTEM; GRAPHFORGE_FILESYSTEM_DURABILITY=strict refuses them. Hosted notebook disks are ephemeral, so export before the runtime recycles.
• Ingest rate. S22 measured 184,749 edges/s, below the 1,000,000 edges/s target. Strong multicore ingest scaling is not established.
• GDC suites. These are failed or incomplete. Unresolved resource, coverage and reference outcomes are tracked in #2178.
• Private repositories and browser-login qualification are outside this release's tested scope.

Full changes

The full change list since v0.5.2 is longer than GitHub's release-body limit; see the comparison and the v0.6.0 milestone.