code-context-control 2.28.0

Local code-intelligence layer for AI coding tools (Claude Code, Codex, Gemini, Copilot). Retrieve less, read less, edit safer.

Code Context Control (C3)

The local code-intelligence layer for AI coding tools.
Stop burning tokens on whole-file reads, blind greps, and unbounded log dumps.
Works with Claude Code, Codex, Gemini CLI, and VS Code Copilot.

---

Why C3?

LLM-driven coding tools have one expensive failure mode: they read too much. They cat whole files, regex the entire repo, dump 10k-line logs into context, and burn through budget before they touch a single line of code.

C3 is a thin, local layer that sits between your IDE and your repo and forces the model to read like a senior engineer:

Without C3	With C3
Read whole 2k-line file	c3_compress → 70% smaller map, then c3_read(symbols=…) for the exact function
Grep the whole repo blindly	c3_search returns ranked candidates with TF-IDF + symbol awareness
Dump full pytest output into the prompt	c3_filter distills 500 lines → 30 actionable ones
Edit, hope it compiled	c3_edit writes via a ledger + c3_validate runs pyright/tsc automatically
Bash test runs that hang on Windows	c3_shell returns structured {exit_code, stdout, stderr, duration} with auto-filter
Lose context on /clear	c3_session(snapshot) + c3_memory persist decisions across sessions

Everything runs locally. No source code or prompts ever leave your machine unless you explicitly opt into a third-party model API.

---

Install

Requires Python 3.10+.

# Clone and install
git clone https://github.com/drknowhow/code-context-control.git
cd code-context-control
pip install .

# Initialize a project (interactive)
c3 init /path/to/your/project

This walks you through:

1. IDE selection (Claude Code / Codex / Gemini / VS Code).
2. Optional local git init.
3. MCP server registration.
4. (Claude Code only) Permission tier selection.

Headless / scripted install:

c3 init /path/to/project --force --ide claude --mcp-mode direct --permissions standard

---

What you get

Surgical reads

c3_compress(file_path="services/router.py", mode="map")    # structural outline (~30% tokens)
c3_read(file_path="services/router.py", symbols=["ModelRouter._classify_features"])

Ranked search

c3_search(query="prompt cache hit rate", action="code", top_k=5)

Safer edits + automatic validation

c3_edit(file_path=..., old_string=..., new_string=..., summary="fix off-by-one")
c3_validate(file_path=...)        # pyright/tsc if installed; syntax check otherwise

Distilled output

c3_filter(text=pytest_output)     # 500 lines → 30 actionable
c3_shell(cmd="pytest -x")          # structured, auto-filtered

Persistent memory

c3_memory(action="recall", query="why did we switch to direct MCP mode")
c3_session(action="snapshot")     # before /clear

A web dashboard for everything you've done

c3-hub                            # http://127.0.0.1:3330

The Hub gives you per-project edit ledgers, session history, memory graphs, benchmark dashboards, and IDE/MCP profile management.

---

Tiered local AI (optional)

C3 ships with optional Ollama integration for offload work the primary model shouldn't waste context on:

Tier	Model class	Used for
Nano	qwen2:0.5b	Sub-100 ms intent routing
Micro	deepseek-r1:1.5b	Last-turn Q&A, summarization
Base	llama3.2:3b+	Code analysis, technical reasoning

c3_delegate(task="summarize this 4k-line stacktrace", backend="ollama")

Ollama is optional. C3 works fine without it.

---

Permissions (Claude Code only)

C3 can manage Claude Code permission tiers via .claude/settings.local.json:

Tier	What it allows
read-only	Exploration — no file writes, no git writes, no installs
standard	Normal dev workflow — edit, build, test, local git (recommended)
permissive	Full trust — everything except destructive ops

All tiers always allow C3 MCP tools and include a deny list that blocks rm -rf, sudo, git push --force, etc.

c3 permissions show
c3 permissions standard

---

Benchmarks

c3 benchmark /path/to/project

Runs C3 against open task suites (Aider Polyglot, SWE-bench Lite) with local-Ollama delegate measurement included.

---

Security & Privacy

C3 runs locally. The Hub binds to 127.0.0.1 by default. Do not expose the Hub to a public network without putting auth/TLS in front of it.

No telemetry by default. Opt-in Sentry crash reporting is available via the [telemetry] extra and requires both SENTRY_DSN (your project) and C3_TELEMETRY_OPT_IN=1 to activate. Even when enabled, request bodies, local variables, and prompts are stripped before sending.

See SECURITY.md for vulnerability reports, the full hardening guide, and the telemetry scrubbing details.

---

License

• Core / OSS — Apache License 2.0 (see LICENSE).
• Pro tier (planned: license-key activation, hosted team Hub, advanced memory) — separate commercial EULA (see EULA-PRO.md).
• Third-party deps — listed in THIRD_PARTY_LICENSES.md.

---

Links

• Changelog: CHANGELOG.md
• Security policy: SECURITY.md
• Issues / questions: open a GitHub issue or email [e-mail]