v0.3.0rc2 — Claude Code setup and Ollama startup

Setup can now connect Codex, Claude Code, or both, and can start an installed local Ollama service after confirmation.

Setup

uvx --from preciso-graphrag==0.3.0rc2 preciso-graphrag setup --project .

Select an agent interactively, or pass --client codex, --client claude, or --client both. Optional provider installation retains that choice. Agent registration uses official client CLIs, shows a preview, preserves unrelated entries, leaves matching entries unchanged, and confirms package-version updates. Claude Code uses private local scope for the selected project.

If Ollama is installed but stopped, selecting Ollama offers to run ollama serve in the background. Setup waits for readiness. A running service is reused. No local service is started for an unavailable remote endpoint. Successful startup leaves Ollama running; failed startup cleans up only the process started by setup. Normal MCP startup remains noninteractive and requires the selected provider to be available.

Existing users

Rerun setup for the same project to retain the model and graph. Stop its active MCP server before verification. An existing managed connection pinned to 0.3.0rc1 can be updated after confirmation. Custom/conflicting configurations are preserved and require manual review.

Both agents share the same collection. Use one agent's MCP server at a time and stop it before switching. This release does not add simultaneous multi-process graph access.

Verification

293 offline tests passed locally. Both real agent CLIs passed isolated registration, idempotency, preservation, and version upgrades. The installed wheel's complete both-agent wizard verified cached Granite and registered both clients using temporary configuration. Real Ollama startup succeeded on a temporary port with model-metadata checks only; no inference or changes to the user's running service. Strict Twine and distribution contents/source-byte checks passed. Hosted checks and TestPyPI/public verification are recorded with the publishing workflow.

Limits

This remains a macOS release-candidate preview. Claude Desktop and other clients require manual registration. Real API credential/Keychain writes and Intel macOS, Windows, and Linux automated onboarding remain unverified. Existing collection model identities and graph formats are unchanged; back up graph data before upgrading. No runtime dependency was added.

TestPyPI upload and all hosted Python, package, and container checks passed: https://github.com/Preciso-GR/preciso-graphrag/actions/runs/39029955703 . The downloaded wheel exactly matched the validated CI artifact and reviewed runtime source.

The TestPyPI-downloaded wheel passed real Codex and Claude Code registration, preservation, idempotency, and version-upgrade checks in isolated client settings. Personal client settings were not modified.

Production publication succeeded: https://pypi.org/project/preciso-graphrag/0.3.0rc2/ . All release checks passed on the first production attempt: https://github.com/Preciso-GR/preciso-graphrag/actions/runs/39030163683 . The public uvx setup command exposed all three client choices. The downloaded public wheel exactly matches production CI and all 42 modules match the reviewed source. Production wheel SHA256: d0aa19cd34f21c13204bf958c1990acc4daac855f37253d86a3d2075890ad090. GitHub assets and SHA256SUMS use those exact production artifacts.

v0.3.0rc1 — guided Codex setup preview

Guided Codex setup preview for individual macOS users. Merged through PR #3 after all hosted checks passed.

Changes

• Interactive local/API embedding selection and project-specific collections.
• Optional pinned IBM Granite Small English R2 embeddings, with an 188.1 MiB artifact download after confirmation. One CPU inference thread and strict 8,192-token input protection.
• Supported Ollama model detection, optional API credentials in macOS Keychain or environment variables, and verified provider identity.
• Tested uvx launcher preparation and confirmed Codex registration without overwriting unrelated entries.
• Clear user documentation, archived earlier plans, and separate manual MCP client examples.
• Lightweight wheel and source distributions; no docs, models, skills, local documents, or graph data bundled.
• TestPyPI/PyPI Trusted Publishing workflows.

Verification

270 tests passed locally and in the locked Linux arm64 container. Hosted Python 3.11, 3.12, and 3.13 jobs, packaging checks, and container checks passed. Local and hosted MCP lifecycle checks passed. Release artifacts below come from the successful production build in https://github.com/Preciso-GR/preciso-graphrag/actions/runs/39026652398 and passed strict Twine and module/content validation.

TestPyPI verification and production publication

Version 0.3.0rc1 is published on TestPyPI: https://test.pypi.org/project/preciso-graphrag/0.3.0rc1/

The wheel downloaded from TestPyPI exactly matched the validated CI artifact (SHA256 5c34cd4295fd208218e778d072a8435b4beafa73d285549a7c8d3abdde6e44ba). Its installed entry point ran outside the source checkout. Real Granite MCP ingestion, restart persistence, retrieval, overlength rejection, rollback, and recovery passed with cached weights and temporary graph data. Real Codex CLI registration preserved an unrelated entry and was idempotent in a temporary Codex home. The personal Codex configuration was not modified.

Version 0.3.0rc1 is now published on PyPI: https://pypi.org/project/preciso-graphrag/0.3.0rc1/

The production upload workflow succeeded. The wheel downloaded from normal PyPI matches the production CI artifact exactly (SHA256 7fe5aed189de472b6200f5cd450da05e1d8f6af996dc6628b7fea9e8749fe0df). All 42 runtime Python modules match the tested TestPyPI wheel. GitHub downloads and SHA256SUMS use the exact production artifacts.

Install this preview with:

uvx --from preciso-graphrag==0.3.0rc1 preciso-graphrag setup --project .

Limits

This is a preview, not broad production readiness. English is the built-in model target. Real API credential/Keychain writes and Intel macOS, Windows, and Linux automated onboarding remain unverified. Other agents use manual MCP configuration. Separate collections do not provide access control. Existing collections keep their configured embedding model; back up data before upgrading.

PRECISO v0.2.0rc1 — GraphRAG MCP

First published release candidate of PRECISO's core-only GraphRAG MCP server.

Included

• Document Contribution Ledger with safe replacement and deletion.
• Graph inspection, node isolation, removal, labels, merges, and disconnections.
• Central ingestion coordination and complete MCP ingestion results.
• Agent-written summaries with freshness checks and atomic graph/vector/queue updates.
• Minimal multi-stage Docker build, offline backups, and operating guides.

Downloads

The wheel is the installable Python package. The tar.gz file is the Python source
distribution. SHA256SUMS contains checksums for both files. Python 3.11 or newer
is required. Install the downloaded wheel with python -m pip install <wheel-file>.
Configure your embedding provider and connect your agent using
python -m preciso_mcp.server. See the repository README for configuration.

GitHub also provides source archives for this tag. This release does not publish
a Docker registry image or a PyPI package.

Verification and limits

252 tests and lint passed locally during packaging. The preceding implementation
also passed 252 tests in the locked Linux arm64 container and offline MCP document
and summary workflows. The packaged wheel passed an MCP handshake, ingestion,
restart persistence, and query check outside the checkout with offline embeddings
and the existing development dependencies.

Hosted GitHub CI for the preceding commit failed before starting any jobs. Its
cause is unresolved. This prerelease does not satisfy the stable-release CI gate.
Representative retrieval quality, large-collection performance, shared-service
access control, and broader platform validation remain open.

Upgrade

Back up your graph directory before upgrading. Do not let older runtimes write
ledger-enabled workspaces. Rebuild legacy graphs from complete, reviewed
extractions before destructive document maintenance. Document replacement and
deletion currently rebuild the retained collection.