Gpp MCP server

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.

What makes it different

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.

• Continuous capture. A high-frequency timeline records every file

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.

• Graphex: versioned, encrypted project knowledge. Architecture,

scoring, compliance-as-code policies enforced at capture/promote/sync, anomaly detection, and per-changeset token/cost attribution.

• Agent governance. First-class agent identity with reputation

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:

Shortened. The full README is on GitHub.