--- generator: "vaultspec-marketing discovery" title: "Vaultspec: Decision-driven harness for coding agents, and humans." description: "Decision-driven harness for coding agents, and humans. vaultspec structures work as #features, with the decision records and documents that ground them." url: "https://vaultspec.neve.md/" component: "site" --- # Decision-driven harness for coding agents, and humans. Vaultspec implements a structured coding workflow focused on #features, decision records, and the documents grounding them. Its rules, skills, and agents guide your coding agent through each step, and its CLI and MCP server scaffold, validate, and link the documents. [Install]() [Read the docs]() [vaultspec-core v0.3.7 · released 13 Oct 2026]() MIT Claude Code, Gemini CLI, Antigravity, Codex ## Grounded, decision-based plans for long-horizon coding work Each feature moves through five steps, from research to review, and every step writes its record to the project's own .vault/ folder. The steps pay off when a feature needs research and a choice between architectural options before code is written. Small fixes don't need the full pipeline. ### #feature tags tie together the pipeline's steps #### Gather evidence .vault/research/ #### Decide direction #### Plan work #### Track the steps #### Verify code ## Continue work from any provider The research, decisions, and plan live in your vault as Markdown files, not in one provider's chat history, so another agent can read them and pick up where the last one stopped. ``` › Can we implement a new calendar module displaying past release dates, with a month and year switcher and filters? ``` Recording deferred. Claude Code ``` › Can we implement a new calendar module displaying past release dates, with a month and year switcher and filters? ``` Recording deferred. Codex ``` › Can we implement a new calendar module displaying past release dates, with a month and year switcher and filters? ``` Recording deferred. Antigravity Recorded CLI excerpts, edited for presentation. ## Documents managed through templates, schemas, and tools for reading and writing Agents write the prose; the tools scaffold each document, validate its metadata and structure, and keep the links between documents intact. Those links give every feature a history of its decisions, execution records, and related features. ### [What plans are in flight?]() See a feature’s plans, finished steps, and next open step before you start or resume work. ### [Is the vault healthy?]() Find broken links, invalid metadata, and structural errors. Add --fix to correct the ones that can be fixed automatically. ### [Where is this implemented?]() Search by what the code does, in plain words. Add --type code to leave documents out. vaultspec-core status search-index ``` $ vaultspec-core status search-index ``` Recording deferred. vaultspec-core vault check all ``` $ vaultspec-core vault check all ``` Recording deferred. vaultspec-rag search "fail the page if it invents customers testimonials or pricing" --type code --limit 1 --scores ``` $ vaultspec-rag search "fail the page if it invents customers testimonials or pricing" --type code --limit 1 --scores ``` Recording deferred. ## vaultspec-core bundles curated rules, skills, and agents you can customize for your project. - ### Rules Tell agents which step comes next and which skill or agent handles it. - ### Skills Walk an agent through each step and the document it produces, from research to review. - ### Agents Specialized personas that research, plan, implement, and review within the framework. - ### Hooks Run conformance and quality checks on the vault documents a session produces. ## Latest changes Each project's current release in full, and the releases before it, taken from the published changelogs when this site was built. ### Current release [0.3.7 13 Oct 2026]() - Fix `ci`make release proposal checks ready without manual approval - Fix `ci`start the proven release when its proposal merges - Performance optimize graph reads and repeated calculations 1. [0.3.5 6 changes 9 Oct 2026]() 2. [0.3.4 6 changes 7 Oct 2026]() [All releases]() ## Install Vaultspec vaultspec-core sets up the harness in your repository. vaultspec-rag is optional, but the workflow is designed around it: it adds semantic search over your code and documents, and installs vaultspec-core with it. ### vaultspec-core Beta Scaffolds and validates the vault Python installs require CPython 3.13 or 3.14 on macOS, Linux, or Windows. With uv, a supported interpreter is downloaded if needed. Standalone binaries need no separate Python installation. From PyPI, with uv or pip Python install method Runs in a cached, isolated environment. Reuse the command prefix for later commands. `uvx vaultspec-core install` Set up a repository & upgrade Install uv for the generated agent launchers, including when you acquire Core through pip or a binary. Use the [command prefix]() that matches your installation. Run vaultspec-core install from your repository root to create the vault and configure your agents. This provisions the repository; it does not install or upgrade the package. After upgrading the package, run vaultspec-core install --upgrade to refresh repository resources. For existing projects, follow the guides to [join a project]() or [upgrade Core](). [Environment settings and API keys]() have their own guide. [Repository]() [Documentation]() ### vaultspec-rag Beta Finds code and decisions by meaning Search runs as a local service. You can use the harness without it. The models need an NVIDIA GPU with CUDA on Linux or Windows, or Apple silicon. One host installation runs the service for the whole machine; a client added to a project installs no GPU packages and uses that service. From PyPI, with uv Installs the command on your PATH for repeated use. `uv tool install --python 3.13 "vaultspec-rag[gpu,mcp]" --index https://download.pytorch.org/whl/cu130 --index-strategy unsafe-first-match` The default host installation: one service for every repository on this machine. On Apple silicon macOS, leave out the two `--index` options. To run the service from one project's own environment instead, follow the [project dependency steps](). Set up a repository & start RAG Install uv for the generated MCP launcher, including when you acquire RAG through Homebrew, Scoop or a binary. After installing the host with any channel above, run this from each repository root to configure its rules, skills and agent connections. The host needs a supported GPU and downloads its models and Qdrant when provisioning. `vaultspec-rag install --no-torch-config` Then start one service for this user on this machine. The service shares the host environment across repositories. Set service environment variables before starting it; exported settings take precedence over saved settings. `vaultspec-rag server start` For a binary, run the extracted executable using its path or add it to your PATH. See the [setup guide](), [upgrade sequence]() and [environment settings guide](). Package upgrades and install --upgrade are separate steps: stop the service, upgrade the host and matching clients, refresh each repository, then start the service once. [Repository]() [Documentation]() After installation, [start a feature with your coding agent]().