MCP server code-context-engine

code-context-engine

Ship the right code, not the whole repo. A local-first MCP context engine for AI coding agents: a per-project code graph, seed resolution from the task as written, and a folded evidence packet — measured on repositories it was never tuned on.

Cursor · Claude Code · Codex · VS Code · Zed · any MCP client

Why · Measured · Install · How it works · Benchmarks · FAQ · Roadmap · Docs

---

Why this fork

NeuroMesh is a strong local-first context engine for web codebases. This repository continues it with two goals that the baseline did not meet:

Goal	Baseline v0.9.0	This project
Project isolation — zero cross-project files in a packet	one live graph per process; a mis-detected workspace could contaminate it	deterministic ProjectId, single-project invariant on the live graph, leakage gate in CI (offered upstream)
Universal codebases — not only web	no notebooks, no config→code layer, no ML node types	.ipynb, YAML/JSON/argparse/Hydra config→code edges, ML artifact graph (Dataset · Model · Checkpoint · Metric · TrainLoop), shell scripts, string-literal index (routes, tool names, env vars)
Honest numbers	one-repo demo table	gold sets on 10 third-party repositories, four of them never tuned on; a model-executed task-success benchmark with a separate judge

Everything the engine claims is in docs/measured.md, with what is not measured listed next to it.

---

What is measured

Gold sets are written from reading the source before the engine runs. Recall = the gold files the packet contains; precision = share of packet files that are gold; forbidden = a file the task must not ship.

Repositories the engine was never tuned on (2026-09-21, main):

Holdout	Languages / stack	Recall	Precision	Forbidden
gin + torchvision	Go, Python	1.000	0.700	0
libuv + fmt	C, C++	1.000	0.589	0
os-lib + cli + Flux.jl	Scala, R, Julia	1.000	0.589	1
peft + keras-hub	Hugging Face / Keras 3 libraries	1.000	0.632	0

Sets that have been tuned on (dev-class): nanoGPT / express / vit-pytorch / fastapi-template 1.000 / 0.938 · django + ultralytics 1.000 / 0.675 · Hydra + argparse configs 1.000 / 0.767 · Fastify + Next.js app router (30 q) 0.800 / 0.608 · this repository (30 real questions) 0.833 / 0.578.

Task success with a real model (DeepSeek-V4-Pro answering, GLM-5.3 judging against the real source, real verify for patch tasks):

Set	packet	whole gold files	naive grep
fixtures (26 tasks)	0.808 · 4.2k tokens	0.808 · 8.7k tokens	0.577
gin holdout (10)	1.000 · 8.1k	1.000 · 11.1k	0.800
torchvision holdout (10)	1.000 · 7.8k	0.800 · 8.7k	0.600

The packet matches or beats "just open the right files" at 1.4–2.1× the success per token, and beats grep everywhere. Recall is the strength; precision (1–3 neighbour files too many) is the open front — see ROADMAP.md.

---

Against the baseline it was forked from

Same 112 gold tasks on 15 third-party repositories, same machine, same day (2026-09-21); NeuroMesh v0.9.0 built from the baseline-v0.9.0 tag, this fork from main. Scored per task by file name from each engine's own optimize output — raw results in docs/baseline-vs-fork-2026-09-21.txt, script in scripts/compare-baseline.sh.

	NeuroMesh v0.9.0 (baseline)	this fork
gold files found (recall, 112 tasks)	0.735	0.938
share of packet files that were wanted (precision)	0.206	0.633
forbidden files shipped	12	7
Scala / R / Julia repositories	0.000 recall (not parsed)	1.000
Hydra + argparse config questions	0.500–0.750 recall	1.000
C / C++ (libuv, fmt)	0.688–0.875 recall	1.000
shell scripts, notebooks, .ps1	skipped at index time	indexed
cross-project leakage	one live graph per process	ProjectId invariant + CI leakage gate

Same recall where the baseline already worked (Go, Python, Django, ultralytics); three times the precision everywhere — the baseline ships four to six files for every one that is wanted. The web-stack front (Fastify + Next.js) is the one place both are weak (0.65 recall), which is why it is the current focus.

What this fork adds

Capability	What it does	Measured by
Project isolation	deterministic ProjectId, single-project invariant on the live graph, a CI test that indexes two projects sharing a file and asserts zero leakage	P0 gate in CI
Config → code	YAML / JSON / .env keys and argparse flags become nodes; cfg.x, args.x, self.hparams.x become Parameterizes edges; Hydra defaults: composition is an import; a key's named reader takes a seat	holdout-cfg 1.000 / 0.712 (F55, F71, F72)
ML artifact graph	Dataset · Model · Layer · TrainLoop · EvalLoop · Checkpoint · Metric nodes from a PyTorch / Keras overlay; notebooks parsed cell by cell	holdout-ml2 1.000 / 0.632 untuned (F26, F58)
String-literal index	quoted names — routes, MCP tool names, event names, env vars, camelCase hooks — resolve to the files that spell them	F75, F78, F80
Twin resolution	same-named definitions settled by directory words, body words and prompt anchors, plural-tolerant; framework-convention stems (route.ts, index, layout) are not "named"	F64, F70, F80, F82
Noise discipline	guesses never land in docs / tests / fixtures / examples unless asked; alias terms must start a word; acronyms resolve exactly	F61, F73, F75
Deterministic packets	every tie has a total order; the same question gives the same packet	F17, F63, F70
A learning loop that learns the path	positive feedback reinforces only the edges a successful task walked; negative feedback reaches required files	F57, F69
Task-success harness	packet vs whole gold files vs grep with a real model, a separate judge that reads the real source, real verify for patch tasks	phase C, F76, F77
12 languages + shell, PowerShell, notebooks	tree-sitter grammars plus light parsers for YAML, JSON, SQL, shell, .ipynb	F78

In a Claude Code session

Same task, same model (Sonnet), claude -p with and without the MCP server, cost from the CLI's own usage report:

Repository	plain (Read / Grep)	with this engine	context tokens	cost
Django, 3,516 files	8 turns · 233k context tokens · $0.239	5 turns · 129k · $0.204	−45%	−15%
fastify/demo, 68 files	8 turns · 244k · $0.408	7 turns · 194k · $0.391	−20%	−4%

The system prompt and tool definitions are re-read every turn and dominate small repositories; the engine only shrinks the code-context share. "99% reduction" figures compare a packet with the whole repository, which no agent sends; the numbers above are what a session actually pays.

---

Install

Pre-built binaries for Linux x86_64, macOS (Apple Silicon) and Windows x86_64 are on the releases page (v1.0.0 is the first).

curl -fsSL https://raw.githubusercontent.com/ParsaVictor/code-context-engine/main/install.sh | sh

irm https://raw.githubusercontent.com/ParsaVictor/code-context-engine/main/install.ps1 | iex

Or from source:

git clone https://github.com/ParsaVictor/code-context-engine
cd code-context-engine
CARGO_BUILD_JOBS=2 cargo build --release -p neuromesh-cli      # target/release/neuromesh

Then, from the root of the project you want to index:

neuromesh doctor                          # binary + workspace check
neuromesh connect --global --agent-rules  # write MCP config + agent rule (Cursor, Claude, VS Code, …)
neuromesh index                           # build the graph (seconds on most repos)

Manual MCP config (any client):

{ "mcpServers": { "neuromesh": { "command": "neuromesh", "args": ["mcp"] } } }

Per-client paths, engines and options: docs/mcp.md · docs/configuration.md.

---

How it works

task as written ──► seed resolution ──► graph activation ──► selection ──► folded packet
                    identifiers, files,   Calls / Imports /    required +   signatures kept,
                    config keys, string   Parameterizes /      callees +    bodies folded,
                    literals, path words  twins, pheromones    ranked fill  budget-capped

1. Index — tree-sitter (and light parsers for YAML/JSON/shell/notebooks) turns every file into nodes and edges; string literals that look like names ("/api/users", "neuromesh_record_feedback", DATABASE_URL) are indexed too.
2. Seed — the prompt's code names resolve to nodes; twins with the same name are settled by directory and body words; guesses that land in docs/tests/fixtures are dropped.
3. Activate & select — energy spreads along typed edges; seeds and the callees the prompt names get required seats; the rest is a ranked, budgeted fill.
4. Fold — bodies outside the question are folded to signatures; the agent can expand a fold on demand.
5. Learn — record_feedback reinforces the path a successful task walked, never a whole neighbourhood.

The MCP agent loop (get_context_packet → search_symbols / expand_gap → expand_fold → trace → record_feedback) is documented in docs/agent-guide.md.

---

Run the benchmarks

bash scripts/benchmark-holdout.sh              # every gold set, one line each (checkouts fetched, graphs cached)
bash scripts/benchmark-holdout.sh holdout-web  # one set
NM_EXPLAIN=1 bash scripts/benchmark-holdout.sh dev   # + per-task reasons in target/explain-dev.txt

Rules the project holds itself to: holdout sets are never tuned on; ratchets only move up; a mixed result is reverted and recorded, never shipped. Every finding since the fork (F1–F82) is written down with its measurement in docs/planning/stage5-findings.fa.md.

---

Documentation

Read this when you want to…
wire an agent to the engine	Agent guide · MCP tools
use the CLI	CLI · Configuration · Engines
understand the design	Architecture · Isolation
check the numbers	Measured · Quality methodology
ask the short questions	FAQ
follow or join the work	ROADMAP.md · Planning & findings (Persian) · Contributing

---

Independent derivative of NeuroMesh by yoosef alipour, MIT. Baseline tag baseline-v0.9.0. See NOTICE and ATTRIBUTION.md. MIT · LICENSE