Apache 2.0 · MCP-native · Private beta
Open-source engineering memory for AI-native dev teams. Built on Postgres + pgvector, served over MCP — bring any LLM, run anywhere your container does.
The cost of forgetting
Across every project your team has shipped, the same fix gets re-derived every month — because no one remembers it's already in last quarter's tickets. The lookup is fast. Only if it happens.
You start a ticket on a site you haven't touched in six months.
You re-read three folders of .md notes and an old PR thread.
Cost: 90 min
You hit a stack trace that feels familiar.
You Slack-search "that 4xx retry loop" across two workspaces.
Cost: 40 min
You ship a fix that someone else shipped six weeks ago.
You'll discover this next Wednesday when the same ticket reopens.
Cost: +2.5 hours rework
→ The knowledge already exists in your team's commits. The recall doesn't. SenterPoint closes that gap in <300 ms per session.
The math, calibrated to research
Most of your team's hardest debugging dies on Slack threads, .md notes no one re-reads, and PR comments archived three quarters ago. SenterPoint targets that specific waste — not vague productivity, not "AI somewhere in your workflow."
≈ $300K / year in fully-loaded engineering time. Per team. Compounding while your corpus grows.
Anchored to Atlassian 2024, McKinsey/IDC, and Stripe/Harris dev-productivity research. Mid-range estimate; upper bound (~50 h/wk) defensible, deliberately left on the table.
How it works
Auto-inject hits A only — sub-300 ms, effectively free. Explicit queries run A + B in parallel, fuse via reciprocal-rank fusion, and hand the shortlist to C. The cheap path runs on every session. The expensive one is opt-in.
Full-text search + structured filter (site / tags / kind).
Runs on every session start to auto-inject the most relevant atoms into the model context — sub-300 ms end to end. No LLM call, no embedding round-trip.
Cosine similarity over 768-d local embeddings.
Runs independently of A, not as a refilter. Both paths return their top candidates; the fusion strip below combines them.
What you get
Each capability maps to a single HTTP endpoint and a single MCP tool with the same shape. No SDK lock-in, no special client library — just JSON over HTTP and the MCP standard your editor already speaks.
You open a site you haven't touched in six months…
A SessionStart hook detects which customer site you opened and auto-injects the top 5–10 atoms before you type a word. You read what your past self learned. No prompt, no recall effort.
GET /auto-inject You're about to write a fix…
Three-stage cascade with citations. The judge filters keyword-only matches and returns 1–3 actually-relevant atoms plus a synthesis pointing at the exact files. The 90 minutes you would have spent re-deriving — gone.
GET /query · query_knowledge() Atom drafts itself from your commit metadata. One keystroke. Zero hand-typed knowledge.
POST /precedents Promote a fix to canonical; the old atom stays searchable, ranks below. No deletes, ever.
POST /precedents/{id}/supersede Tenant-scoped, site-blind. Ask once; recall spans every site you own.
GET /atoms?tags=… Same business logic behind every surface. OpenAPI generated from running code — cannot lie.
/openapi.json · /mcp Speaks MCP — works where you already write code
Built like infrastructure
Postgres for everything (FTS, structured filters, embeddings). No exotic vector store, no orchestration layer, no abstraction tax. The whole service is one process — what runs in your mesh is what runs in our tests.
Private beta · 30+ devs · 2 design partners · 50+ sites in production
Single container plus Postgres 16 + pgvector. Bearer-token auth, Caddy reverse proxy, runs anywhere your container does — no managed-service hooks, no upstream phone-home. Apache 2.0.
your editor ─MCP─▶ senterpoint ──▶ postgres + pgvector ┊ ┊ ┊ └─── your mesh / your VPC / your hardware ───┘
Public source follows the beta cohort, once we've finished the git-history audit. Design partners can request a tarball today.
Stage C speaks the OpenAI API shape — any provider, any local endpoint (llama.cpp, vLLM, Ollama), any model. Rotate via env var. Your key, your spend, your data path.
Atoms scoped per-tenant at the data layer (composite PK, every query filtered by tenant_id). Multi-tenant isolation is structural, not a WHERE clause we hope to remember.
Built on open source · upstream credit where it's due
Ready when you are
Apache 2.0. Model-agnostic. Private beta — tell us about your team and we'll wire the cascade for you.