Sophon

Deterministic context compression for MCP agents. One Rust binary. Zero ML at query time. Reproducible benchmarks, real-data measurements.

npm version npm total downloads GitHub release CI License: MIT Rust 1.75+ MCP Tests

Sophon is a deterministic context layer for agents speaking the Model Context Protocol. It compresses prompts, conversation memory, code digests, file deltas, and shell output — without an embedding model at query time, without a GPU, and without API keys.

Single 5.2 MB Rust binary. MCP-native. cl100k_base-accurate. Default build pulls no Python, no ML weights, no network.


What it does, in 30 seconds

Tool What it solves
compress_prompt Long structured prompt → keep only sections relevant to the query
compress_history Growing conversation → summary + facts + recent window + optional retrieval
compress_output Shell stdout/stderr → 21 domain-aware filters (git, cargo, docker, kubectl, JSON, …)
read_file_delta / write_file_delta Re-reads + edits → diffs only, never the whole file
encode_fragments Repeated boilerplate → single token reference
update_memory Append turn → JSONL persist + incremental rolling summary
navigate_codebase Repo digest with tree-sitter / regex + PageRank, ranked by query

11 MCP tools total (full table below).


Real numbers — measured on this repo's own dev cycle

We built four independent benches that each capture a different chunk of an agent's tool traffic. All four run against this repo's actual git history + working tree on the operator's machine. Reproducible byte-for-byte by anyone with cargo build --release.

Dimension What it measures Saved Bench
history compress_history over real commits 94.6 % real_session_capture.py
shell compress_output on real git/cargo/gh/ls stdout 84.4 % real_session_shell.py
filereads compress_prompt on real Rust / Python / Markdown / TOML files 71.7 % real_session_filereads.py
search compress_output on real grep/find patterns 79.5 % real_session_search.py
🎯 Weighted blend (35/30/20/15) typical agent session estimate 84.7 % real_session_holistic.py

real_session_holistic.py runs all four sub-benches with --json, parses them, and produces the weighted blend. Default weights reflect this repo's observed shape; pass --weights "history=0.4,..." to model your own workload.

USD economy on Claude Opus 4.7

Saved per session
Naive input pricing ($15/MT) $2.03
With prompt caching (25-turn reads at $1.50/MT) $3.24

Pass --model sonnet or --model haiku to real_session_deep_dive.py if you're re-pricing for a cheaper tier.

Where each dimension falls short (we say it ourselves)

  • history measures only what git captures (commits + diffs) — typically ~5-10 % of a real session's tool traffic. The 94.6 % is the upper bound, not the typical case.
  • shell mixes commands that compress well (git diff 95 %) with commands that don't (gh repo view --json adds tokens, −9 %). 84.4 % is a real-world average, not a curated highlight.
  • filereads uncovered that compress_prompt on raw source files compresses by budget cap, not by query routing — same file with 3 different queries → identical output. Section detection only fires on structured input (Markdown headers, XML tags). Documented inline in the bench.
  • search depends entirely on YOUR repo's state. A repo with no TODOs gets 0 % on grep TODO.

The blended 84.7 % is napkin-math from a linear weighted average across four real measurements. Not a cherry-picked synthetic. Run the benches yourself to verify.

Other reproducible benchmarks (synthetic, on-thesis)

Test Result Bench
compress_output across 18 command families 90.1 % weighted aggregate compress_output_per_command.py
25-turn synthetic Claude Code session 68.1 % session tokens saved session_token_economics.py
compress_prompt across 22 prompt shapes 70.2 % mean, 36 ms mean latency prompt_compression_extended.py
Code retrieval on "where is X?" questions recall@3 = 70 % (vs grep 10 %, FULL 20 %) repo_qa.py
vs LLMLingua-2 on structured prompts +8.9 pt accuracy at 35× lower latency llmlingua_compare.py
Sophon + Anthropic prompt caching +24 % tokens / +49 % $ on top of caching sophon_plus_prompt_caching.py
Sophon + mem0 Additional savings on retrieved memories sophon_plus_mem0.py

Why Sophon — "in front of X"

Sophon is not a memory platform, a recall system, an OCR stack, or a replacement for provider-side caching. It's a deterministic compressor that slots in front of whatever memory / cache / code-nav layer you already use, and attacks the tokens those layers can't.

In front of Anthropic / OpenAI prompt caching

Provider caching handles the static half of a request — system prompt, tool definitions, reused documents. It doesn't touch the dynamic half (growing conversation history, tool outputs). Sophon compresses exactly that half. The two stack cleanly.

+24 % tokens / +49 % $ saved on top of prompt caching on a 25-turn Claude session — because the uncached dynamic block is billed at 10× the cached rate. See sophon_plus_prompt_caching.py.

In front of mem0 / Letta / Zep / Graphiti

Memory systems retrieve the right memories. Sophon shrinks what gets sent to the LLM after retrieval. If mem0 returns 2 kB of raw memories, compress_prompt keeps only the sections the query actually references.

Honest caveat: on very short retrieved blocks (< ~200 tokens) Sophon's wrapper adds overhead and you should pass through. The bench reports this directly.

In front of Claude Code / Cursor / Cline

Primary use case. Every repeat file read becomes a read_file_delta; every shell command output goes through compress_output; every repeated boilerplate block gets a fragment_cache token. Install transparently with sophon hook install --agent claude --global.

In front of a RAG pipeline

navigate_codebase produces a PageRanked repo digest that a RAG retriever would otherwise spend expensive embedding calls to build. Tree-sitter / regex symbol extraction over 11 languages, sub-second.

When NOT to use Sophon

  • Long-form conversational recall above 80 % — Sophon caps at ~40 % on LOCOMO and we don't chase it. Run mem0 / Letta / Zep for recall, then optionally pipe their output through Sophon.
  • Multi-hop reasoning on massive documents — that's HippoRAG or GraphRAG.
  • OCR / PDF layout — out of scope. Use Docling / Marker / Unstructured upstream.
  • Very small inputs (< ~200 tokens) — Sophon's section scaffolding can cost more than it saves.

Quick start

Install via npm (recommended)

npm install -g mcp-sophon
sophon doctor          # verify install + show config

The postinstall script downloads the right prebuilt binary for your platform from the GitHub Releases page. Supported: macOS arm64/x64, Linux arm64/x64, Windows x64.

Build from source

git clone https://github.com/lacausecrypto/mcp-sophon
cd mcp-sophon/sophon
cargo build --release -p mcp-integration       # ~5.2 MB binary

Optional features:

# 11-language tree-sitter AST extraction (~25 MB):
cargo build --release -p mcp-integration --features codebase-navigator/tree-sitter

# BGE-small semantic embedder (~34 MB), activate with SOPHON_EMBEDDER=bge:
cargo build --release -p mcp-integration --features bge

# All features (~42 MB):
cargo build --release -p mcp-integration --features "codebase-navigator/tree-sitter,bge"

Requires Rust 1.75+.

Wire it into an MCP client

Most clients accept this snippet (Claude Desktop, Claude Code, Cursor, Cline, Continue):

{
  "mcpServers": {
    "sophon": {
      "command": "sophon",
      "args": ["serve"]
    }
  }
}

Run sophon doctor to print the right config path for your client.

Recommended runtime setup

# Persistent memory + on-disk retriever store + BM25+Hash hybrid
export SOPHON_MEMORY_PATH=~/.sophon/memory.jsonl
export SOPHON_RETRIEVER_PATH=~/.sophon/retriever
export SOPHON_HYBRID=1

sophon serve

Quick CLI

sophon exec -- cargo test                       # run + compress combined output
sophon compress-prompt --prompt ./system.txt --query "rust errors" --max-tokens 500
sophon hook install --agent claude --global     # transparent Claude Code integration
sophon stats --period session                   # token savings rollup

Programmatic (one-shot JSON-RPC)

echo '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"compress_prompt","arguments":{"prompt":"<rust>?: operator</rust><web>fetch()</web>","query":"rust errors","max_tokens":500}}}' \
  | sophon serve

What the binary ships

11 MCP tools, all stdio:

Tool What it does
compress_prompt Keep query-relevant sections of a long prompt
compress_history Summary + facts + recent + optional retrieval over the conversation
compress_output Strip noise from command stdout/stderr (21 domain filters + JsonStructural)
navigate_codebase tree-sitter / regex digest of a repo, PageRanked by query
update_memory Append messages, JSONL persist, optional rolling summary
read_file_delta Version/hash-aware file read, unchanged → minimal payload
write_file_delta Send edits as diffs, not full files
encode_fragments / decode_fragments Detect repeated boilerplate, swap with tokens
count_tokens cl100k_base-accurate token count
get_token_stats Session-level savings rollup

Binary sizes by feature set:

Build Size
Default (regex extractors, HashEmbedder) 5.2 MB
+ tree-sitter (11 languages) ~25 MB
+ BGE semantic embedder ~34 MB
All features ~42 MB

MCP protocol: 2025-06-18. notifications/cancelled actually drops the response (since v0.5.4). Structured JSON-RPC error codes (-32000..-32099 reserved for Sophon). Infallible dispatcher — a malformed request can't kill the stdio loop.


Configuration

Run sophon doctor to see every SOPHON_* env var currently set with validation warnings. Full catalogue (24 flags) lives in runtime_flags.rs. The flags worth knowing:

Flag Effect Cost
SOPHON_RETRIEVER_PATH=/dir Activate the semantic retriever (chunk store on disk) ~0
SOPHON_MEMORY_PATH=/file.jsonl Persistent conversation memory across sophon serve runs ~0
SOPHON_HYBRID=1 BM25 sparse-lexical + HashEmbedder fused via RRF ~1 ms
SOPHON_ROLLING_SUMMARY=1 Build rolling summary at update_memory time, not at query time LLM call moved to ingest
SOPHON_CHUNK_TARGET=500 Bigger chunks preserve cross-sentence context ~0
SOPHON_EMBEDDER=bge Swap HashEmbedder for BGE-small (needs --features bge) model load at startup
SOPHON_LLM_CMD="claude -p --model haiku" LLM shell-out command (used by summarizer when configured) per-call subprocess

Deprecated v0.4.0 recall-chasing flagsSOPHON_HYDE, SOPHON_FACT_CARDS, SOPHON_ENTITY_GRAPH, SOPHON_ADAPTIVE, SOPHON_LLM_RERANK, SOPHON_TAIL_SUMMARY, SOPHON_REACT, SOPHON_GRAPH_MEMORY, SOPHON_MULTIHOP_LLM — chase LOCOMO recall, an axis we no longer optimise. Still functional but sophon doctor flags them. Removed in a future major.


Honest limitations

The full list lives in BENCHMARK.md § 8. Headlines:

  • LOCOMO conversational recall caps at ~40 %. mem0 / HippoRAG hit 80-90 % with neural retrieval at query time — we chose determinism + sub-100 ms p99 instead. Pipe mem0 in front of Sophon if you need that recall.
  • HashEmbedder is keyword-bound. "favorite food" ↔ "weakness for ginger snaps" doesn't match. Activate BGE (SOPHON_EMBEDDER=bge) for semantic recall — costs +25 MB binary + model load.
  • No multimodal ingestion. Images / PDFs / audio out of scope. Run Docling / Marker / Unstructured upstream.
  • Rolling summary doesn't help on small sessions. When the un-summarised tail fits the budget, the rolling cache is a no-op. Useful for long-running sessions with SOPHON_LLM_CMD set.
  • Some commands don't compress. gh repo view --json adds tokens, git log --oneline saves 0.4 %. Sophon's job isn't to compress already-compact output — it's to compress redundant verbose output. The benches name the gaps explicitly.

Project layout

.
├── README.md           ← you are here
├── BENCHMARK.md        ← full per-section benchmark detail
├── CHANGELOG.md        ← version history + deprecated numbers
├── benchmarks/         ← reproducible scripts for every number above
├── npm/                ← npm wrapper package
└── sophon/crates/      ← 11-crate Rust workspace
    ├── prompt-compressor/    compress_prompt
    ├── memory-manager/       compress_history, update_memory, rolling summary
    ├── delta-streamer/       read/write_file_delta
    ├── fragment-cache/       encode/decode_fragments
    ├── semantic-retriever/   chunker + HashEmbedder + BM25 + entity graph
    ├── output-compressor/    21 command-aware filters + JsonStructural
    ├── codebase-navigator/   tree-sitter / regex + PageRank
    ├── cli-hooks/            transparent agent installer
    └── mcp-integration/      stdio server, async dispatch, cancellation

Contributing

PRs welcome. Run the test suite:

cd sophon && cargo test --workspace --lib --tests --exclude prompt-compressor   # 405 tests
cd sophon && cargo test --features codebase-navigator/tree-sitter               # +AST tests
cd sophon-py && .venv/bin/pytest tests/                                         # 4 Python tests

Every benchmark claim is reproducible — pointers to the scripts live in BENCHMARK.md. If a number doesn't reproduce on your machine, open an issue.

Particularly welcome:

  • TypeScript bindings (Python bindings ship in sophon-py/)
  • gh family filter (gh run list, gh pr list, gh repo view --json) — the bench shows this is currently a gap
  • SOPHON_EMBEDDER_CMD shell-out plugin pattern (mirror of SOPHON_LLM_CMD) for Voyage / OpenAI / Cohere
  • Multi-repo real_session_holistic.py runs against popular open-source repos

License

MIT. See LICENSE.