🗺️ Cartograph

Build a personal cognitive graph of everything you work with — and plug it into your AI agents.

Point it at your folders. It maps your repos, notes, and docs into one searchable graph, learns your field, and serves the right context to Claude Code, Cursor, or any MCP agent — so every future task, from coding to daily generative-AI use, is grounded in your knowledge.

Local-first · your data never leaves your machine · works on any field · no ML expertise required

Quickstart · How it works · Connect your agents · Requirements · Install tiers · Visual app

---

✨ Why

LLM agents are brilliant but amnesiac — they don't know your repos, your conventions, your past decisions, or what "good" looks like in your field. Cartograph fixes that locally:

1. Plug — pip install cartograph__v1
2. Ingest — carto ingest ~/code builds a graph of your work (incrementally, any folder)
3. (Optional) Train — add semantic search / your own models on your graph
4. Auto-applied — your agents query the graph over MCP for grounded context, forever

No cloud. No account. Your graph is a single SQLite file you own.

---

🚀 Quickstart

pip install cartograph__v1            # PyPI package name; the command is `carto`, the import is `cartograph`
carto demo                        # ⚡ see it ALL work in ~10s on a synthetic corpus (zero setup)
carto init                        # pick folder(s) + your field(s) — declaring your field makes labels accurate
carto ingest                      # build your graph (re-run anytime; only changed files reprocess)
carto viz                         # 👀 see your graph in the browser
carto retrieve "how did I handle auth" --chunks   # hybrid search over everything you've done

Want semantic (meaning-based) search too?

pip install "cartograph__v1[semantic]"   # ~2GB; uses your GPU if present, else CPU
carto index                          # embed your chunks once

---

🧠 How it works

flowchart TB
    subgraph YOU["Your stuff (stays on your machine)"]
        R["repos · notes · docs · books"]
    end
    subgraph CARTO["Cartograph"]
        ING["Ingest: walk → chunk → infer field (incremental)"]
        GRAPH["The graph — SQLite + FTS5\nprojects · files · chunks · skills · edges"]
        SEM["Semantic index (optional)\nlocal embeddings, GPU-aware"]
        HY["Hybrid retrieval (RRF of semantic + keyword)"]
        EL["Elite layer\nfrontier catalog · playbooks · Definition-of-Done · coverage"]
    end
    subgraph OUT["Where it shows up"]
        MCP["MCP server → Claude Code / Cursor / any agent"]
        VIZ["Desktop visual graph"]
        CLI["carto CLI"]
    end
    R --> ING --> GRAPH
    GRAPH --> SEM --> HY
    GRAPH --> HY --> EL
    HY --> MCP & VIZ & CLI
    EL --> MCP & CLI

Loading

The graph schema — everything you work with becomes connected nodes:

flowchart LR
    Project -->|CONTAINS| File -->|CHUNKED_INTO| Chunk
    Project -->|IN_FIELD| Field[("field: ml / quant / web / …")]
    Project -->|RELATED_TO| Project
    Project -.->|REPO_TEACHES| Skill
    Chunk -->|FTS5 + embeddings| Search[("hybrid retrieval")]

Loading

Hybrid retrieval is the headline. It fuses meaning-based (semantic) and exact (keyword) search via Reciprocal Rank Fusion. In the engine Cartograph distills from, hybrid scored success@10 0.986 vs 0.958 for either method alone — it's never worse, because it catches both paraphrases (semantic) and exact tokens like a function name (keyword). It works in pure-keyword mode with zero ML installed, and turns on semantic automatically once you run carto index.

The elite layer pulls any build toward the top of its field (works for ML, quant, web, HPC, data, devops, mobile, game-dev, research, libraries — and is one file to extend):

• carto elevate "<task>" — the elite bar, reference repos, the frontier playbook (process), and the repos you already have to build on
• carto frontier — how much of your field's top-tier reference set you've ingested + what to add
• carto review <project> --field <f> — grade a build against the field's Definition-of-Done

🧭 The persona layer — steer your agents to you

The graph knows your work; the persona layer learns what you respond to and shapes every answer. It models you as field weights + an optional preference vector in embedding space, re-ranks retrieval by alignment to you, and emits a model-agnostic steering brief any agent prepends — so Claude / Cursor / ChatGPT / Gemini outputs adapt to your field, conventions, and preferences, and keep adapting as you give feedback. Confidence-scaled: well-supported preferences steer hard, sparse ones barely nudge.

carto persona                         # your learned focus + confidence
carto personalize "how do I cache this?"   # the steering brief an agent prepends
carto feedback --liked my-repo        # teach it what was useful (adapts over time)

The same brief is available to agents via the MCP personalize tool and to web GenAI via carto serve + a tiny userscript (docs/BROWSER.md). Foundations, the Hilbert-space mapping, and honest limits: docs/PERSONA.md.

---

🔌 Connect your agents

Cartograph speaks MCP (Model Context Protocol). Add it once and your agent can query your graph.

One-step wiring: run carto agent-setup — it prints the MCP config and a system-prompt rule that makes your agent call Cartograph automatically each task (personalize → retrieve_context → record_use).

Claude Code / Cursor — add to your MCP config (~/.cursor/mcp.json or Claude Code's MCP settings):

{
  "mcpServers": {
    "cartograph": { "command": "carto", "args": ["mcp-server"], "type": "stdio" }
  }
}

Your agent now has these tools:

Tool	What it gives the agent
personalize	call first — a steering brief (your persona, field, output guidance + your relevant snippets) so the answer fits you and adapts over time
retrieve_context	the relevant code/doc snippets to inject (hybrid)
relevant_projects	which of your repos relate to the task
elevate_task	top-of-field briefing: bar + reference repos + playbook
frontier_status	your coverage of each field's best references
record_use	after answering, report what helped → the persona adapts automatically
graph_stats	size of your graph

Tip: tell your agent in its system prompt "At the start of a task, call elevate_task and retrieve_context against Cartograph." — then every future task is grounded in your knowledge.

---

💻 Requirements

	Minimum	Recommended
Python	3.10+	3.12
OS	Windows / macOS / Linux	any
RAM	4 GB	16 GB+
Disk	~50 MB + your data	put the graph on a fast/large drive (CARTOGRAPH_HOME)
GPU	none (CPU works)	any CUDA GPU → ~10× faster embedding
Heavy ML	not required	cartograph__v1[semantic] for meaning-based search

Efficiency tips

• Put your workspace on a fast, roomy drive: export CARTOGRAPH_HOME=/mnt/fast/cartograph.
• A CUDA GPU makes carto index dramatically faster; CPU still works (just slower).
• carto ingest is incremental — re-run it anytime; only changed files reprocess.
• Brute-force semantic search is fine to a few million chunks (~150 ms/query). Past that, see docs/SCALING.md to swap in FAISS.

---

📦 Install tiers & graph sizes

Cartograph is modular — install only what you need, and grow the graph to any size:

Install	Command	Adds
Core	pip install cartograph__v1	full graph + keyword search + viz + Studio + MCP. Tiny, instant.
Semantic	pip install "cartograph__v1[semantic]"	meaning-based + hybrid search; the learned models (~2 GB)
ML	pip install "cartograph__v1[ml]"	train your own graph models on your data
Vision	pip install "cartograph__v1[vision]"	real-time screen capture → OCR → graph (carto watch; needs Tesseract)
Secure	pip install "cartograph__v1[secure]"	at-rest encryption for sensitive data (e.g. screen captures)
Everything	pip install "cartograph__v1[full]"	all of the above

Graph-size tiers — don't download one giant graph. Choose what fits:

• Your own (recommended): carto ingest your folders — the graph is exactly your scale.
• Starter reference packs (optional, public OSS only — never anyone's personal data): pre-built graphs of curated top-tier repos per field, offered as quantized download tiers (S / M / L) via GitHub Releases. Pick a small pack to seed a new field, or build your own with scripts/build_reference_pack.py. See docs/REFERENCE_PACKS.md.

---

🖥️ The visual app

For non-technical users, one command opens an interactive map of your knowledge:

carto viz        # or double-click scripts/launch_viz.bat (Windows) / launch_viz.sh (mac/linux)

• Pan/zoom a force-directed graph of your projects, colored by field.
• Type a query → relevant projects light up and matching snippets appear.
• Zero setup, runs locally in your browser, no data leaves your machine.

---

🔒 Privacy

• Local-first. Everything lives in ~/.cartograph (or CARTOGRAPH_HOME). Nothing is uploaded.
• Your data is git-ignored by default; the graph, index, and config never get committed.
• Reference packs contain only public OSS — never personal data.

---

🧩 Extending

Add your field in three small files and Cartograph elevates it like any other: cartograph/elite/catalog.py (reference repos) · playbooks.py (the process) · dod.py (the bar). Field inference lives in cartograph/ingest.py.

🗂️ Commands

carto demo · agent-setup · init · ingest · index · train · route · watch · retrieve · elevate · frontier · review · persona · personalize · feedback · prefs · serve · studio · secure · stats · viz · mcp-server · doctor (run carto --help).

License

MIT — see LICENSE. Use it, fork it, build your own.