io.github.omkar9854/token_optimizer

slimctx — the token optimizer for AI agents

Zero-dependency, fully-reversible context compression for AI agents.

slimctx compresses what your agent reads — tool outputs, logs, JSON, source files, prose — before it reaches the LLM. Same answers, fraction of the tokens. Pure Python stdlib: no ML models, no downloads, no network calls, ever. Auditable end to end in ~1,600 lines.

Live output of python3 benchmarks/demo.py — run it yourself, nothing is staged.

from slimctx import Pipeline, Config

pipe = Pipeline(Config(target_tokens=32_000))
result = pipe.compress(messages)        # OpenAI/Anthropic-style dicts
print(result.savings_ratio)             # e.g. 0.82

original = pipe.retrieve("a1b2c3d4...")  # byte-exact original, any time

Results (synthetic workloads modeled on real agent traffic)

Every run also verifies that each planted "needle" (the FIXME, the OOMKill, the outlier) survives compression, and that every lossy transform is byte-exact reversible. Reproduce with python3 benchmarks/bench.py.

How it works

messages ──► ContentRouter ──► one of:
                ├─ JSON  : lossless tabularization (repeated keys → header,
                │          constant columns → legend), then relevance-ranked
                │          row selection only if still over budget
                ├─ LOG   : Drain-style template mining — repeated lines
                │          collapse to `pattern [x1432]`; errors verbatim
                ├─ CODE  : AST skeleton — signatures + docstrings kept,
                │          bodies elided EXCEPT those relevant to the query
                └─ TEXT  : extractive sentence selection (BM25 + salience
                           + position), verbatim, never paraphrased

The four guarantees

1. Universal reversibility. Before any lossy transform, the original goes into a content-addressed store (memory / SQLite / bring-your-own cipher) and the output carries a [slimctx-ref <hash> ...] marker. The model — or you — can always get the byte-exact original back.
2. Errors are never dropped. Every compressor pins error/warning content: log errors pass verbatim, salient JSON rows are kept, salient sentences outrank filler.
3. Deterministic output. Same input → byte-identical output, across runs and processes. Compressed prefixes stay stable, so provider prompt-caches (Anthropic/OpenAI) keep hitting.
4. Net gain or no-op. If a transform doesn't save enough tokens to pay for its marker, the original is kept untouched. The live zone (system prompt + last N messages) is never modified at all.

Why not just use Headroom?

Headroom is the established project in this space and is more featureful today (provider proxy with SSE streaming, agent wrappers, cross-agent memory, an ML compression model). slimctx makes a different set of trade-offs, aimed at locked-down / client-site deployments:

If you need the proxy/wrap ecosystem, use Headroom. If you need something you can read in an afternoon, run air-gapped, and certify for a client environment, use slimctx.

Install / test

pip install slimctx           # from PyPI — or vendor the slimctx/ directory
python -m pytest tests/ -q    # 26 tests: invariants, not examples
python3 benchmarks/bench.py   # reproduce the numbers above

Integration sketches

As a library (any framework): call pipe.compress(messages) right before your provider SDK call; expose pipe.retrieve as a tool named retrieve so the model can pull originals.

As an MCP server (GitHub Copilot, Claude Code, Cursor, ...): ships built in, stdlib-only:

python3 -m slimctx.mcp_server --db ~/.slimctx/store.db

See USAGE.md for the GitHub Copilot (.vscode/mcp.json) setup and a security deployment checklist.

Encrypted store:

from cryptography.fernet import Fernet          # optional, your choice
f = Fernet(key)
store = SqliteStore("ccr.db", cipher=(f.encrypt, f.decrypt))
pipe = Pipeline(store=store)

License

Apache-2.0. Original implementation — no code derived from Headroom.