Skip to content
pmLet’s talk
All projects

Nella

Codebase intelligence for AI coding agents.

Nella sits between an AI coding agent and your repo. Instead of letting the model guess from training data, it indexes the real codebase: AST-chunked, with a semantic vector index and a BM25 lexical index built from a single chunk pass. The agent searches actual code and gets back file paths, line ranges, and confidence scores. On top of retrieval, Nella keeps a persistent memory: typed assumptions that auto-invalidate when the file they depend on changes, a change ledger, and dependency-drift detection, all surviving across sessions in a local .nella/ directory.

Nella feature page with code indexing and search examples on an angled MacBook
Role
Product design and engineering
Year
2026
Platform
Software project
Status
live

The problem

LLM coding agents don't have a memory problem so much as a grounding problem: they invent imports, forget decisions they made five turns ago, and act on stale assumptions about a schema that changed an hour ago. Bigger context windows just let them hallucinate more confidently.

What I built

The work includes @getnella/mcp, @usenella/core, @usenella/api.

The choices behind it.

Solo · design + engineering

AST-chunked hybrid index

A semantic vector index (Voyage voyage-code-3) and a BM25 lexical index build from one chunk pass, with hybrid, semantic, and lexical query modes plus per-result confidence scores. The agent searches the real repo instead of guessing from training data.

Implementation choice

Persistent memory across sessions: a typed assumption ledger that auto-invalidates when a watched file changes, a change ledger, and dependency-drift snapshots over package.json and the lockfile. Catches contradictions before broken code ships.

Multi-agent coordination over MCP

Parallel agents register, claim tasks, record decisions, and get conflict checks against shared state. A challenge-response heartbeat verifies session continuity between tool calls.

Prompt-injection backstop

An injection scorer rates each indexed chunk for risk and flags suspicious repo content, so hidden instructions in code or comments can't quietly hijack the agent.

Implementation choice

Companion benchmark suite grades agents (Claude Sonnet, Opus, GPT-4o) on capability and safety: build and test pass, constraint violations, scope creep, and cost. It retries failed tasks with error feedback and renders an HTML comparison dashboard.

Three surfaces over one core: an MCP server, a CLI, and the embeddable @usenella/core library

Agent-agnostic, deployed on GCP Cloud Run and Cloud SQL, with a hosted MCP at mcp.getnella.dev.

Nella: public website
Public website
Build detail
@getnella/mcp v0.2.7 published to npm
Build detail
4-package pnpm workspaces monorepo (core, MCP/CLI, REST API, benchmark)
TypeScript source files, 57 test files
225
Build detail
Rate limiting: 2 algorithms (sliding-window, token-bucket) x 3 backends (memory, Redis, SQLite)

Under the surface.

@getnella/mcp

mcp+cli

The published package (v0.2.7): an MCP server plus the nella CLI that indexes a repo, then hands grounded search, code context, and multi-agent coordination tools to Claude Code, Cursor, and VS Code.

@usenella/core

library

The engine behind everything else: AST-based chunking, hybrid semantic and BM25 search, a dependency graph, and persistent context that tracks assumptions, a change ledger, and dependency drift.

@usenella/api

service

The Express REST service exposing workspace, search, validation, context, and auth endpoints, backed by a BullMQ job queue and WebSocket progress updates.

@usenella/benchmark

library

A suite that runs coding agents (Claude Sonnet/Opus, GPT-4o) against tasks and scores pass rate, constraint violations, scope creep, and cost, plus a prompt-injection detection test across eight attack categories.

Build details Stack, scope, and implementation

Built with

  • TypeScript
  • Node.js
  • pnpm workspaces
  • Model Context Protocol (MCP)
  • Express
  • usearch (vector index)
  • BM25 via natural
  • @typescript-eslint/typescript-estree (AST)
  • Voyage AI voyage-code-3 embeddings
  • PostgreSQL / Cloud SQL
  • Supabase Auth
  • Redis + BullMQ
  • GCP Cloud Run + Cloud Storage
  • Zod

Areas

  • MCP
  • AI coding agents
  • RAG
  • code intelligence
  • dev tool
  • multi-agent

More about the project

It ships as three surfaces over one core: an MCP server, a CLI, and an embeddable library. That means it plugs into Claude Code, Cursor, or any MCP-compatible agent without lock-in. Parallel agents can register, claim tasks, and record decisions against shared state, with a challenge-response heartbeat keeping the trust chain intact. A companion benchmark package red-teams agents on capability and safety, including multi-turn prompt-injection scenarios. The hosted MCP runs on GCP Cloud Run behind mcp.getnella.dev.

Product animation