# How Heimdall Token Saver Compares

Heimdall Token Saver focuses specifically on **command output compression** — it doesn't cache, delegate, or summarize via LLM. The tools below solve adjacent but different problems, and in many cases can be used alongside Heimdall Token Saver.

## Comparison Table

| Feature | Heimdall Token Saver | cc_token_saver_mcp | token-optimizer-mcp | Claude Context Mode |
|---------|------------|-------------------|--------------------|--------------------|
| **Approach** | Deterministic compression per command type | Delegates simple tasks to local LLM | Caching + compression via MCP | Sandboxed execution + FTS5 indexing |
| **Requires LLM calls** | No | Yes (local LLM) | Yes | No |
| **Added latency** | ~0ms (regex/parsing only) | Variable (LLM inference) | Variable | ~0ms for sandbox, variable for indexing |
| **Platform support** | Claude Code, Antigravity CLI | Claude Code only | Claude Code only | Claude Code only |
| **Compression method** | 36 specialized processors (git, pytest, cargo, go, terraform, pulumi, docker, k8s...) | Task delegation (not output compression) | Response caching | Output sandboxing + summarization |
| **Preserves all errors/traces** | Yes (precision-tested) | N/A | Depends on cache hit | Depends on summary |
| **Works offline** | Yes | Needs local LLM running | No | Yes |
| **Install complexity** | `python3 install.py` | MCP server config | MCP server config | MCP skill install |

## Key Differences

### Heimdall Token Saver

Heimdall Token Saver intercepts command output and applies **deterministic, per-command compression** using 36 specialized processors. It understands the structure of `git diff`, `pytest`, `terraform plan`, and other common CLI outputs, and removes only noise (progress bars, passing tests, installation logs) while preserving all actionable information (errors, diffs, warnings).

- Zero latency overhead (regex and string parsing only)
- Fully deterministic — same input always produces same output
- Works offline, no external dependencies
- 853 tests including precision tests that verify critical data survives compression
- Supports both Claude Code and Antigravity CLI

### cc_token_saver_mcp

An MCP server that intercepts tool calls and delegates simple tasks (like reading files or listing directories) to a local LLM instead of Claude. This reduces token usage by avoiding Claude entirely for trivial operations. It's a different strategy — **task delegation** rather than output compression. Can be used alongside Heimdall Token Saver.

### token-optimizer-mcp

An MCP-based tool that caches and compresses responses. Useful for repetitive queries where the same command is run multiple times. Complements Heimdall Token Saver's single-pass compression with cross-invocation caching.

### Claude Context Mode

A skill for Claude Code that runs commands in a sandboxed environment and indexes outputs for later retrieval via FTS5. Useful for managing very large codebases where full context doesn't fit. Solves a different problem — **context management** rather than output compression.

## Can I Use Them Together?

Yes. Heimdall Token Saver operates at the output level (compressing what the model sees), while the other tools operate at the task level (delegating), caching level (avoiding re-computation), or context level (indexing). They are complementary, not competing.