Full-lifecycle sovereign Rust development assistant with interactive ReAct agent loop (reedline), 5 core sovereign tools (read, patch, write, bash, search_files), Hyper local token proxy & cost detector (:8000), 100% BYOK execution, and zero telemetry.
CACode never hallucinates blind rewrites. It plans sequentially, inspects with 5 sovereign tools, and meters every token locally through port 8000.
internal/middleware/auth.rs to locate token verification handler.<= instead of strict expiry check).patch_file tool.cargo test test_expired_token via sandboxed shell to verify zero regression.read_file(path="internal/middleware/auth.rs", offset=40, limit=25)patch_file(path="internal/middleware/auth.rs", old="if exp < now", new="if exp <= now { return Err(AuthError::Expired); }")execute_command(command="cargo test test_expired_token -- --nocapture")~/.cacode/costs.db (TX #1042)Benchmarked on Linux x86_64 Fedora 44, Rust 1.85, 120k LOC multi-crate workspace.
| Evaluation Metric | CACode (v0.3.0) | Claude Code (Node.js) | Cursor (Electron) | Aider (Python) | Windsurf (Electron) |
|---|---|---|---|---|---|
| Cold-Start Launch | < 15ms (Native Rust) | ~1,200ms | ~2,500ms | ~1,800ms | > 3,200ms |
| RAM Usage (Idle) | < 25 MB | ~120 MB | ~350 MB | ~95 MB | > 450 MB |
| Proxy Interception Latency | 0ms (Embedded Hyper :8000) | N/A (No proxy) | N/A (Proprietary) | N/A (Direct) | N/A (Proprietary) |
| Code & Query Telemetry | 0 bytes (Zero Telemetry) | Opt-out only | Continuous telemetry | Telemetry opt-in | Continuous telemetry |
| Offline Local Models (Ollama/vLLM) | 100% First-Class (BYOK) | Cloud-locked (Anthropic only) | Limited / Cloud-first | Supported | Cloud-locked |
| Autonomous Sub-Agent Concurrency | Built-in Parallel Fleet | Single thread | Single thread | Single thread | Single thread |
Engineered for software engineers who require maximum AI autonomy without giving up code confidentiality or suffering bill shock.
Multiline terminal REPL powered by reedline with syntax highlighting, smart completion, and autonomous Reason + Act planning that self-corrects upon test failures.
High-performance native primitives: windowed read_file, AST fuzzy patch_file, embedded ripgrep search_files, sandboxed execute_command, and tree-sitter inspect_symbols.
Embedded Hyper reverse proxy on localhost:8000 intercepting LLM inferences, accelerating prompt caching, metering live token consumption, and enforcing hard session budgets.
Every token consumed and dollar spent is recorded to an immutable local SQLite database (~/.cacode/costs.db) with zero external ledger reporting.
Bring your own keys for OpenAI, Anthropic, Ollama, vLLM, Groq, and 9Router. Your API keys and source code never leave your local environment.
Orchestrate parallel sub-agent workers for simultaneous bug fixing, test coverage repairs, and documentation generation in isolated workspaces.
Rapid engineering workflows via /plan, /review, /test, /diff, /compact, /model, /undo, /cost, and /settings directly from the prompt bar.
All file edits, unified diff applications, and terminal script executions run inside verified security sandboxes with explicit user approval gates.
Engineered in pure Rust 2024. Cold launch in under 15 milliseconds and idle memory consumption under 25 MB RAM—10x lighter than Node or Python agents.
Structural AST validation of git diffs before committing, guaranteeing zero hallucinated imports or breaking API regressions enter your repository.
Seamlessly switch across multiple repositories and branch contexts with persistent session journals that restore instantly without loss of context.
Fully compatible with offline local models via Ollama and vLLM. Zero telemetry, zero analytics tracking, and complete code ownership in your hands.
100% BYOK. Bring your own keys. Pay once for permanent pro features with zero predatory subscription lock-in.
Perfect for individual developers and researchers wanting a sovereign terminal coding assistant with local BYOK models.
For power engineers and professionals who require sub-agent concurrency, prompt cache optimization, and lifetime commercial licensing.
For engineering leads and teams requiring unlimited concurrent sub-agent fleets, custom tool extensions, and proxy clustering.
Everything you need to know about code privacy, ReAct loop, and the :8000 cost proxy.
CACode operates under strict 100% BYOK and Zero-Telemetry principles. All configuration (~/.cacode/config.toml) and cost ledgers (~/.cacode/costs.db) reside exclusively on your local disk. No telemetry, prompt logs, or source code snippets are ever transmitted to Fathforce servers.
CACode's sovereign primitives include: read_file (windowed offset file reader), patch_file (fuzzy AST-aware unified diffing that avoids destructive rewrites), search_files (embedded ripgrep regex and file search), execute_command (sandboxed terminal runner with user confirmation gates), and inspect_symbols (tree-sitter symbol outline navigator).
CACode runs an ultra-lightweight Hyper reverse proxy on localhost:8000. Every inference call is metered in real-time to compute token counts, prompt cache hits, and exact dollar costs recorded into local SQLite. You can set hard budget caps (e.g. abort if session spend exceeds $0.50) to prevent runaway loops.
Absolutely. CACode features native integration with Ollama, vLLM, and LM Studio. Simply point your provider to your local server (e.g. http://localhost:11434 for Qwen 2.5 Coder or DeepSeek) and CACode will function seamlessly with zero internet access required.