Code Context Engine MCP server

---

Use cases

	Use case	How CCE helps
💰	Reduce Claude Code costs	94% fewer input tokens per session
🔒	Keep code private	Everything local, no cloud indexing
🔄	Multi-editor teams	One index across Claude Code, Cursor, VS Code, Gemini CLI
🧠	Cross-session memory	Decisions and context survive restarts
⚡	Faster responses	Less context = faster Claude replies
📊	Track actual savings	Dollar amounts, not estimates

---

Quick start

One command. 30 seconds.

uvx --from "code-context-engine[local]" cce init    # install + index + configure, one shot

Or if you prefer a persistent install:

uv tool install "code-context-engine[local]"    # or: pipx install "code-context-engine[local]"
cd /path/to/your/project
cce init

Restart your editor. Done. Every question now hits the index instead of re-reading files.

Agent Plugin support: Run cce init --plugin to generate a portable Agent Plugin directory that works with VS Code, Cursor, Copilot, Codex, ChatGPT, and Kiro. The plugin uses uvx to launch CCE on demand, so users don't need to pre-install the Python package. See Agent Plugin below.

Already have Ollama? Skip [local] and use uv tool install code-context-engine instead. CCE auto-detects Ollama at localhost:11434 and uses nomic-embed-text.

Python 3.11+ and a C compiler (for tree-sitter grammars).

Platform	Setup
macOS	xcode-select --install
Ubuntu/Debian	sudo apt install build-essential cmake
Fedora/RHEL	sudo dnf install gcc gcc-c++ cmake
Windows	Visual Studio Build Tools (C++ workload) + CMake

Tested on macOS, Linux, Windows with Python 3.11/3.12/3.13.

cce init auto-detects your editor and writes the right config. To target a specific agent, use --agent claude, --agent codex, --agent copilot, --agent pi, or --agent all.

Editor	Config written	Instructions
Claude Code	.mcp.json	CLAUDE.md
VS Code / Copilot	.vscode/mcp.json	.github/copilot-instructions.md
Cursor	.cursor/mcp.json	.cursorrules
Gemini CLI	.gemini/settings.json	GEMINI.md
OpenAI Codex	~/.codex/config.toml (user-global, per-project section)	AGENTS.md
OpenCode	opencode.json
Tabnine	.tabnine/agent/settings.json	TABNINE.md
Pi	.mcp.json	AGENTS.md

Multiple editors in the same project? All get configured in one command.

Codex note: Codex CLI reads MCP servers from ~/.codex/config.toml only — it has no per-project config. cce init adds one [mcp_servers.cce-<project>-<hash>] section per project so multiple projects coexist; cce uninstall removes only the section for the current project.

Pi note: Pi does not support MCP natively. To use CCE with Pi, you need a pi MCP adapter extension (e.g. pi-mcp-adapter) that consumes the .mcp.json config and exposes CCE's tools to the Pi agent. cce init sets up both .mcp.json and AGENTS.md (Pi loads the latter automatically for startup instructions).

  my-project · 38 queries · last query 5m ago

  ⛁ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶  88% tokens saved

  Input savings   1.9M  tokens   $27.78
  Output savings  4.8k  tokens   $0.36
  ──────────────────────────────────────────
  Total saved   1.9M  tokens   $28.15

  Breakdown:
    retrieval              84%  ▰▰▰▰▰▰▰▰▰▰    1.8M   $26.76 · 12 calls
    chunk compression       3%  ▰▱▱▱▱▱▱▱▱▱   68.5k    $1.03 · 12 calls
    output compression*    <1%  ▰▱▱▱▱▱▱▱▱▱    4.8k    $0.36 · 12 calls

  Cost estimate based on Opus pricing (input $15.0/1M, output $75.0/1M)

Supports Anthropic, OpenAI, and Google model pricing. Configure via pricing.model in ~/.cce/config.yaml.

---

Why this matters

Input tokens are 85-95% of your Claude Code bill. CCE cuts them by 94% (benchmarked on FastAPI).

Without CCE:    Claude reads payments.py + shipping.py   = 45,000 tokens
With CCE:       context_search "payment flow"            =    800 tokens

	Without CCE	With CCE
Session startup	Re-reads files every time	Queries the index
Finding a function	Read entire 800-line file	Get the 40-line function
Cross-session memory	None	Decisions + code areas persisted
Token cost (Sonnet, medium project)	~$0.14/session	~$0.04/session

---

Benchmark: FastAPI (reproducible)

We benchmarked CCE against FastAPI (53 source files, 180K tokens) with 20 real coding questions. No cherry-picking, no synthetic queries.

Methodology: For each query, "without CCE" means reading the full content of every file the query touches. "With CCE" means the relevant chunks after compression.

Important baseline note: The 94% number is measured against full-file reads, not against what Claude Code actually does. In practice, Claude Code already uses grep, partial file reads, and targeted tools, so the real-world savings compared to normal Claude Code behavior will be lower than 94%. We use full-file as the baseline because it's reproducible and deterministic (no agent behavior variability). The benchmark measures CCE's retrieval efficiency, not a head-to-head comparison with Claude Code's built-in exploration.

Metric	Result
Retrieval savings	94% (83,681 → 4,927 tokens/query)
Compression (additional, on retrieved chunks)	89% (4,927 → 523 tokens/query)
Recall@10 (found the right files)	0.90
Latency p50	0.4ms
Queries tested	20

Per-Layer Savings (each measured independently)

Layer	What it does	Savings	Method
Retrieval	Full files → relevant code chunks	94%	measured
Chunk Compression	Raw chunks → signatures + docstrings	89%	measured
Grammar	Drops articles/fillers from memory text	13%	measured

Output compression (reducing Claude's reply length) provides additional savings (~65% estimated) but is not included in the headline number above.

Multi-language benchmarks

Repo	Language	Files	Retrieval savings	Recall@10
FastAPI	Python	53	94%	0.90
Django	Python (large)	2,347	93%	0.95
Express	JavaScript	6	94%	1.00
chi	Go	94	76%	0.67
fiber	Go (monorepo)	396	93%	0.07

Django (2,347 files, 5.4M tokens) shows CCE scales to large codebases with 0.95 recall. Go's shorter files reduce the retrieval headroom (smaller baseline). Monorepos dilute recall at top-10 (fiber). Middleware queries with one-feature-per-file hit R=1.00 consistently.

Reproduce it yourself:

pip install code-context-engine
python benchmarks/run_benchmark.py --repo https://github.com/fastapi/fastapi.git --source-dir fastapi
python benchmarks/run_benchmark.py --repo https://github.com/go-chi/chi.git --source-dir .

Full results in benchmarks/results/. Queries and methodology in benchmarks/.

---

What you get

11 MCP tools that Claude uses automatically:

Tool	What it does
context_search	Hybrid vector + BM25 search with graph expansion
expand_chunk	Full source for a compressed result
related_context	Find code via graph edges (calls, imports)
session_recall	Recall decisions from past sessions
session_timeline	Walk turn summaries for a session (drill into recall hits)
session_event	Inspect raw tool input/output for a specific event
record_decision	Save a decision for future sessions
record_code_area	Record which files were worked in
index_status	Check index freshness
reindex	Re-index a file or the full project
set_output_compression	Adjust response verbosity (off / lite / standard / max)

Live dashboard with donut charts, file health, and session history:

cce dashboard

Dollar estimates with multi-provider pricing (Anthropic, OpenAI, Google):

cce savings --all    # see savings across all projects

---

How it works

1. Index: Tree-sitter parses your code into semantic chunks (functions, classes, modules). Stored as vector embeddings locally.
2. Search: Claude calls context_search. Hybrid vector + BM25 retrieval finds the right chunks. Code graph adds related files automatically.
3. Compress: Chunks are truncated to signatures + docstrings (or LLM-summarized if Ollama is running).
4. Remember: Decisions and code areas persist across sessions via session_recall.
5. Track: Every query is logged. cce savings shows exactly how much you saved.

Re-indexing after edits takes under 1 second (96% embedding cache hit rate). Git hooks keep the index current automatically.

---

What makes CCE different

It saves where the money is

Output compression tools (like Caveman) save 20-75% on output tokens. Output is 5-15% of your bill. Net savings: ~11%.

CCE saves on input tokens (94% retrieval savings on FastAPI, reproducibly benchmarked). Input is 85-95% of your bill.

It actually understands your code

Not a text search. Tree-sitter AST parsing creates semantic chunks. Hybrid retrieval merges vector similarity with BM25 keyword matching via Reciprocal Rank Fusion. A confidence scorer blends similarity (50%), keyword match (30%), and recency (20%). Graph expansion walks CALLS/IMPORTS edges to pull in related code.

It remembers

record_decision("use JWT for auth", reason="session tokens flagged by legal") is stored in SQLite and surfaces via session_recall in the next session. No re-explaining your architecture.

It tracks real savings

Not estimates. Actual tokens served vs full-file baseline, broken down by buckets (retrieval, compression, output, memory, grammar). Dollar costs fetched from Anthropic's pricing page. Savings summary shown at every session start.

It is secure by default

Secret files (.env, *.pem, credentials.json) are never indexed. Content is scanned for AWS keys, GitHub tokens, Slack tokens, Stripe keys, JWTs, and generic credentials. PII (emails, IPs, SSNs, credit cards) is scrubbed from memory writes. All MCP file paths are validated against path traversal.

---

Under the hood

SHA-256 fingerprint per chunk, salted with model name. Re-index skips unchanged code. Binary float32 storage (10x smaller than JSON). Typical re-index: 96% cache hit, under 1 second.

Replaced LanceDB with sqlite-vec. Same cosine-distance quality, 99% smaller install. WAL mode + PRAGMA NORMAL for 80% write speedup. Vectors, FTS5, code graph, and compression cache all in three SQLite files.

Memory entries compressed without LLM calls. Drops articles, fillers, pronouns. Three levels (lite/full/ultra, 20-60% savings). Code, paths, URLs preserved byte-for-byte. Same input always yields same output.

5 Claude Code lifecycle hooks capture session context. Every hook runs curl ... || true, so a crashed server never blocks the user. SessionStart injects bootstrap context; others capture silently.

Dollar estimates in cce savings support 15+ models across Anthropic, OpenAI, and Google. Static pricing ships with CCE, live Anthropic pricing is fetched and cached 7 days. Configure pricing.model (e.g. gpt-4o, gemini-2.5-pro, sonnet) or override with pricing.input / pricing.output for custom rates.

Running dozens of cce serve processes (one per project per AI session) can exhaust system memory. The resource governor caps ONNX Runtime threads per process, uses advisory file locks so only one process indexes a given project at a time, backs off under Linux memory pressure (PSI), and auto-shuts down idle servers after 30 minutes. Configure via serve.idle_timeout_minutes and serve.max_ort_threads.

CCE's cross-session memory depends on the agent calling record_decision and record_code_area. Memory nudges make recording ambient: after N searches without a recording, context_search results include a short reminder. At session end, the Stop hook summarizes unrecorded activity. Nudges re-arm after the first recording so they stay useful without being noisy.

cce serve --http exposes a POST /search endpoint for custom agent integrations that speak HTTP instead of MCP stdio. Same hybrid retrieval pipeline, structured JSON response with confidence scores. Input validation clamps top_k (1..100) and confidence_threshold (0.0..1.0).

7 buckets track every token saved: retrieval, chunk compression, output compression, memory recall, grammar, turn summarization, progressive disclosure. Survives restarts. Powers CLI and dashboard analytics.

---

Agent Plugin

Agent Plugins is an open standard (v1.0.0) backed by Amazon, Cursor, Microsoft, OpenAI, and Vercel for packaging AI skills and MCP servers into portable, zero-install bundles. CCE can generate a plugin directory that compatible editors can discover and load automatically.

cce init --plugin                          # Generate at .cce/plugin/
cce init --plugin --plugin-dir ~/plugins/cce  # Custom location
cce init --agent claude --plugin           # Both: agent config + plugin

What gets generated

.cce/plugin/
├── plugin.json                       # Agent Plugins v1.0.0 manifest
├── mcp.json                          # MCP server config (uvx + stdio)
├── skills/
│   └── code-context/
│       ├── SKILL.md                  # Agent instructions (frontmatter + body)
│       └── references/
│           └── tools.md              # Per-tool parameter docs (loaded on demand)
└── LICENSE

Compatible editors

VS Code, GitHub Copilot, ChatGPT, Codex, Cursor, and Kiro. The plugin uses uvx to launch CCE on demand, so users do not need to pre-install the Python package. The MCP server auto-discovers the project root by walking up from its working directory, looking for .context-engine.yaml or .git/.

When to use --plugin vs --agent

	--agent (default)	--plugin
Install method	Writes editor-specific config files	Generates a portable plugin directory
Zero-install	No, CCE must be on PATH	Yes, uvx fetches CCE on demand
Instruction updates	Stale until cce init re-run	Stale until cce init --plugin re-run
Best for	Your own machine	Sharing with a team or distributing

Both can be used together. --agent handles per-editor MCP config, --plugin provides a portable alternative.

---

CLI at a glance

cce init                    # Index + install hooks + register MCP
cce init --plugin           # Generate Agent Plugin for VS Code, Cursor, etc.
cce                         # Status banner
cce savings                 # Token savings with dollar estimates
cce savings --all           # All projects
cce dashboard               # Web dashboard with live charts
cce search "auth flow"      # Test a query
cce status                  # Index health + config
cce services                # Ollama + dashboard + MCP status
cce commands add-rule '...' # Project rules for Claude
cce uninstall               # Clean removal of all CCE artifacts

Run cce list for the full command reference.

---

Configuration

Zero-config by default. Override what you need in ~/.cce/config.yaml or .context-engine.yaml:

compression:
  level: standard          # minimal | standard | full
  output: standard         # off | lite | standard | max
  ollama_url: http://localhost:11434   # point at a remote Ollama if desired

retrieval:
  top_k: 20
  confidence_threshold: 0.5

pricing:
  model: opus              # opus | sonnet | haiku | gpt-4o | gemini-2.5-pro | ...
  # input: 15.0            # override $/1M input tokens
  # output: 75.0           # override $/1M output tokens

Remote Ollama: If you run Ollama on another machine in your network, set compression.ollama_url (e.g. http://nas.local:11434) or export CCE_OLLAMA_URL (the env var wins). CCE probes the endpoint and falls back to truncation-only compression when it's unreachable, so a flaky link won't break indexing.

---

Output Compression

CCE also compresses Claude's responses (same concept as Caveman):

Level	Style	Savings
off	Full output	0%
lite	No filler or hedging	~30%
standard	Fragments, drop articles	~65%
max	Telegraphic	~75%

Tell Claude: "switch to max compression" or "turn off compression". Code blocks and commands are never compressed.

---

Disk Footprint

Component	Size
Core install (Ollama backend)	~17 MB
With [local] extra (fastembed + ONNX)	~189 MB
Embedding model (one-time download)	~60 MB (fastembed) or managed by Ollama
Index per project (small/medium/large)	5-60 MB

No GPU required. With Ollama, embeddings are handled by the Ollama server. With the [local] extra, the embedding model runs on CPU via ONNX Runtime.

---

Supported Languages

AST-aware chunking (tree-sitter parsed, 11 extensions):

Language	Extensions
Python	.py
JavaScript	.js, .jsx
TypeScript	.ts, .tsx
PHP	.php
Go	.go
Rust	.rs
Java	.java
C#	.cs

Language-aware fallback chunking (40+ extensions):

Category	Languages
Web	HTML, CSS, SCSS, LESS, Vue, Svelte
Systems	C, C++, Zig, Nim
Mobile	Swift, Kotlin, Dart
Functional	Haskell, Scala, Clojure, Elixir, Erlang, F#
Scripting	Ruby, Perl, Lua, R, Bash/Zsh
Data/Config	JSON, YAML, TOML, XML, SQL, GraphQL, Protobuf
DevOps	Terraform, HCL, Dockerfile
Docs	Markdown

All other text files are chunked by line range. Binary files are skipped.

---