gpp (git++)

AI agents change code continuously, then lose the work that happened between commits — and the notes they keep about your codebase (CLAUDE.md, memory banks, knowledge files) go stale silently when the code moves on. gpp is a version control system built for that reality: it captures every change as it happens, promotes curated changesets with intent and provenance, and hosts project knowledge on the repo's own history — so when a commit invalidates something you believed about the code, gpp can name that commit.

(Recording is generated by scripts/demo.sh — deterministic, reproducible, no mockups.)

30-second try

cargo install gpp-cli

gpp init --graphex
echo "fn main() {}" > main.rs
gpp timeline                     # captured already — no staging, no commit
gpp promote -m "first cut" --intent feature
gpp diff HEAD                    # semantic diff: renames/moves are one op

And the part no other VCS does — record what you believe about the code, and let history police it:

gpp belief add --claim "token expiry is 24h" --evidence src/auth.rs:7-7
# ...weeks of commits later...
gpp belief bisect "token expiry is 24h"
# INVALIDATED  cs:fhcpef7c  "raise token expiry to 7 days"
#  -     7 | pub const EXPIRY_HOURS: u64 = 24;
#  +     7 | pub const EXPIRY_HOURS: u64 = 168;

Deterministic — diff intersection and blob hashes, zero LLM or network calls. (For scale: asked to judge whether a stored memory has been invalidated, the best frontier model manages 55.2% accuracy on the STALE benchmark — here it's a history query, not a judgment.) Validated on real history across five repos in four languages — axum 0.6→0.7, flask 1.1→2.0, clap 3→4, zod 3→4, go-redis 8→9 — where every invalidated belief bisects to a pinned, documented culprit commit and the control beliefs survive: see demos/belief-bisect/ for the full matrix.

Related MCP server: ITHZ MCP

What makes it different

• Continuous capture. A high-frequency timeline records every file change (SQLite WAL, debounced watcher); curated changesets are promoted from it with an explicit intent, author kind (human/agent), and cost record. Nothing is lost between commits.

• Graphex: versioned, encrypted project knowledge. Architecture, conventions, decisions, and beliefs live in an encrypted knowledge graph inside the repo, tier-gated per agent trust level, audited on every read — and staleness-checked against the history it rides on.

• Agent governance. First-class agent identity with reputation scoring, compliance-as-code policies enforced at capture/promote/sync, anomaly detection, and per-changeset token/cost attribution.

Everything else — P2P sync over Noise, replay, review/RBAC, a relay node, a TUI — is the platform underneath: see docs/ARCHITECTURE.md.

Git is the substrate, not the competition. The bridge (gpp git-import / git-export / git-bridge) round-trips real Git commits, so GitHub and existing workflows keep working — gpp's knowledge, provenance, and governance layers ride alongside. The axum demo above runs entirely on history imported through this bridge.

Connect Claude Code (or any MCP client)

gpp ships an MCP server; agents query the knowledge graph (every belief with a freshness envelope — anchor, commits since, and the changeset that staled it), write their own evidence-anchored beliefs with propose_belief (human-approved, then policed by history), propose changesets, and report costs. Drop this in .mcp.json at your repo root:

{
  "mcpServers": {
    "gpp": { "command": "gpp", "args": ["mcp-server", "--stdio"] }
  }
}

Full client setup (Claude Desktop, generic stdio) and the exposed tool list: docs/MCP.md.

Status

All 9 roadmap phases (0–8) implemented. See docs/ROADMAP.md for per-phase deliverables and documented deviations, docs/TODO.md for the prioritized backlog, and docs/WORKLOG.md for the running engineering log.

Verified 2026-08-23: 184 workspace tests pass, cargo clippy and cargo fmt clean, full workspace builds. No stub crates remain — every crate has a working implementation. Coverage is measured in CI (cargo llvm-cov; 65.7% line at baseline, being raised).

Test depth is still uneven: foundational layers are well covered (gpp-core, gpp-graphex, gpp-diff, gpp-tui ≥ 80% line; the CLI has end-to-end suites for policy enforcement, cost reporting, reviewer assignment and belief bisect), while several integration crates remain at smoke-level (gpp-sdk, gpp-notify, gpp-rbac, gpp-replay). "Implemented" here means built and tested against its milestone, not exhaustively hardened everywhere. Closing that gap is the top item in docs/TODO.md.

The full layer table

Also: extensions/{gh-gpp,vscode-gpp,neovim-gpp}, GitHub Actions + GitLab CI templates, deploy/ Docker images, packaging/ Homebrew.

Documented follow-ups (recorded in the ROADMAP/TODO, not silently skipped): registry/license APIs for deps, native PyO3/napi bindings, outbound platform-review sync, apt/dpkg packages.

Install / build

cargo install gpp-cli                    # the `gpp` binary (crates.io)
cargo install gpp-relay                  # relay node (optional)

# or from a clone
cargo build --release
cargo test --workspace
cargo bench -p gpp-core -p gpp-diff      # criterion perf suite

Prebuilt binaries for Linux, macOS (ARM + Intel), and Windows are attached to each release; cargo install --git https://github.com/mahabubul470/gpp gpp-cli tracks unreleased development.

More to try

# Governance
gpp policy template secrets-scan
gpp trust show
gpp audit --include-cost --include-graphex

# Decentralized: sync two repos over Noise
gpp sync serve 127.0.0.1:9473             # on peer A
gpp sync add a 127.0.0.1:9473 && gpp sync # on peer B

# GitHub-compatible
gpp remote setup --platform github --repository acme/webapp
gpp remote pr-create --base main

See CLAUDE.md for project context, docs/ for the full specification (architecture, data model, CLI, protocols, roadmap), and docs/book/ for the user guide + tutorials.

License

MIT