codepulse

Adaptive runtime intelligence for AI coding agents: live behavioral model + static structure (including indexer-backed structural search), queried through a compact MCP surface.

Status: implementable MVP — Rust daemon, MCP server, Python + .NET agents, demos and E2E scripts.

Site: detroitpro.github.io/codepulse · alpine-realm-ney2.here.now (static landing in site/)

Why

Agents today mostly see source text. codepulse adds what actually runs: call counts, hot paths, observed edges, joined with AST-derived complexity — plus structural pattern search via the tree-sitter indexer (ast-grep-like questions over MCP, no separate tool) — with an adaptive probe controller so exact instrumentation is time-boxed and budgeted.

Example structural asks: “find all async defs”, “where is requests.get called?”, “show bare except:”, “find open(...) sites”, “locate execute methods” — then join to runtime tools. Pattern table: docs/MCP_API.md.

Stack

Piece	Choice
Core (ingest, store, controller, indexer)	Rust (axum + SQLite)
Static parsing + structural search	tree-sitter (Python + C#)
MCP server	TypeScript (@modelcontextprotocol/sdk)
Runtime agents	Python ≥3.12 (sys.monitoring) · .NET (Lib.Harmony)

The platform language is not dictated by the target runtime. Runtime agents are plugins that speak a shared agent protocol.

Quickstart

# 1) Daemon (indexes --root, listens on :7420)
cargo run -p codepulse-daemon -- --root examples/python-demo --db .codepulse/dev.db

# 2) MCP server (another terminal)
cd packages/mcp && npm install && npm run dev
# point your AI host at this stdio server; CODEPULSE_ENDPOINT=http://127.0.0.1:7420

# 3) Python demo under the agent
pip install -e "agents/python[dev]"
CODEPULSE_ROOT=$PWD/examples/python-demo CODEPULSE_ENDPOINT=http://127.0.0.1:7420 \
  python -c "from codepulse_agent import install; install(); import uvicorn; from app import app" \
  # or use scripts/e2e-python.sh

E2E:

chmod +x scripts/*.sh examples/dotnet-demo/scenario.sh
./scripts/e2e-python.sh
./scripts/e2e-dotnet.sh

Repo layout

codepulse/
  docs/                 Design docs + user guide (docs/guide/)
  site/                 Landing + static user docs (GitHub Pages)
  skills/codepulse/     Agent Skill (SKILL.md)
  crates/               Rust workspace (daemon + libraries)
  packages/mcp/         TypeScript MCP server
  agents/python/        CPython agent
  agents/dotnet/        .NET agent (Harmony)
  examples/             python-demo + dotnet-demo
  scripts/              e2e-python.sh / e2e-dotnet.sh

Agent Skill: skills/codepulse — teach agents when to use runtime tools and how to set up the stack.

Docs

User guide

Doc	Contents
docs/guide/GETTING_STARTED.md	Clone, daemon, Python demo, MCP, first hot paths
docs/guide/ADVANCED.md	Structural search, probes, .NET, env, troubleshooting
docs/guide/AGENTS.md	MCP config, tool map, copy-paste bootstrap prompt

Site: Getting started · Advanced · Agents

Design

Doc	Contents
docs/PRODUCT.md	Thesis, personas, wedge, non-goals
docs/ARCHITECTURE.md	Components, flows, overhead/privacy
docs/DATA_MODEL.md	SQLite schema, identity rules
docs/MCP_API.md	Tool contracts
docs/AGENT_PROTOCOL.md	Agent ↔ daemon wire contract
docs/ROADMAP.md	Phase 0–3

Principles

1. Answers over streams
2. Adaptive cost
3. Static + dynamic together (catalog + structural search in the indexer)
4. Local-first
5. Fail closed on privacy
6. Runtime agents are plugins (pattern search is not on the agent wire)

License

MIT