slimtoken

🔀 A token-optimization layer that sits between an Anthropic-compatible client and its backend — a local llama-server or a cloud API — and rewrites every request to use fewer tokens before forwarding it.

Tool schemas are trimmed, the system prompt is compressed, old turns are distilled, repeated tool results are collapsed, and tool output is type-compressed. Fewer input tokens → faster prompt-eval, lower cost, more context headroom. It also ships a backend optimizer that recommends llama-server arguments tuned to your GPU.

You can use slimtoken three ways — all driven by the same core pipeline, none reimplemented:

MIT-licensed. Ships with orjson, xxhash, and tiktoken for fast JSON, hashing, and real token counting. The full pipeline runs on by default — tools, system, messages, dedup, distill, and lossy tool-result compression. Disable any stage with a SLIMTOKEN_* env switch; SLIMTOKEN_MINIFY=0 for raw passthrough. Only output filtering (token cap / stop sequences) is opt-in.

Install

pip install slimtoken

# Proxy: wire ANTHROPIC_BASE_URL at the proxy, then run it
slimtoken install
slimtoken serve --upstream http://127.0.0.1:8082        # local llama-server
slimtoken serve --upstream https://api.anthropic.com   # or cloud

# CLI: minify a request body on demand
slimtoken optimize -i request.json                     # minify a request body
slimtoken presets --measure                            # local-model table + measured reduction

# MCP server: stdio JSON-RPC for MCP clients
slimtoken-mcp                                          # or: python -m slimtoken.mcp_server

Uninstall the proxy wiring: slimtoken uninstall (or bash scripts/uninstall.sh).

slimtoken install only writes a marker block to your shell rc that exports ANTHROPIC_BASE_URL (prior value backed up to ~/.slimtoken/prev_env and restored on uninstall). It never touches settings.json, CLAUDE.md, or mcp.json, so removal is clean and fully reversible.

Related MCP server: everything-slim

How it works — the proxy optimization stack

A six-stage minify pipeline runs on each request, all on by default. The diagram shows the request lifecycle with the t0–t4 latency boundaries the proxy records per request — proxy-side work (ingress + optimize) is what slimtoken controls; model-side (forward → first token → final token) is where real time goes.

sequenceDiagram
    participant C as client
    participant P as slimtoken proxy
    participant B as backend / model
    C->>P: POST /v1/messages  (t0)
    P->>P: read request  (t0→t1)
    P->>P: minify: tools · system · messages · dedup · distill · budget  (t1→t2)
    P->>B: forward minified body  (t2)
    B-->>P: first output token  (t3)
    P-->>C: stream raw bytes back  (t3→t4)
    Note over P: proxy-side = (t1-t0)+(t2-t1) ≈ 12 ms<br/>model-side = (t3-t2)+(t4-t3) — dominates

Safety guarantees — code fences (``` / ~~~) preserved byte-identical; pruning is pair-safe (a tool_result is never orphaned from its tool_use); identity-based change detection returns unchanged content zero-copy; the grammar field is stripped from request bodies.

Input optimization (total)

Total input-token reduction, the always-on config (full pipeline), measured by the pipeline itself on representative payloads — not asserted:

Run the measurement yourself:

slimtoken presets --measure                    # recompute the table above on your machine
python3 bench/benchmark.py                     # payload + latency breakdown
python3 bench/benchmark.py --backend http://127.0.0.1:8082   # + end-to-end vs a live backend
slimtoken latency                             # one request → t0-t4 printout

Proxy latency is ~12 ms per request (optimize stage, warm) — negligible next to any LLM round-trip. The win is fewer tokens sent, not proxy speed. The t0–t4 instrumentation separates slimtoken's work from model generation so you can see exactly that.

One config, no profiles

There are no named profiles. slimtoken always runs the full pipeline (the old aggressive preset, minus the name); every stage and knob is a raw SLIMTOKEN_* env switch. The two things you might actually want to do:

• Turn it all off — SLIMTOKEN_MINIFY=0 (raw passthrough; for debugging or when the model must see input verbatim).

• Turn off one lossy stage — e.g. SLIMTOKEN_MINIFY_DISTILL=0 to keep old turns verbatim, or SLIMTOKEN_TOOL_COMPRESS=0 to keep tool results verbatim.

See the Config table for the full knob list. The single config surface (build_config) is shared by the proxy, CLI, MCP server, and skill.

Backends — Anthropic, OpenAI, and Ollama

The proxy routes by URL path and the CLI/MCP accept a --format / format arg. The minify pipeline is built around Anthropic's request shape; OpenAI and Ollama bodies are normalized to that canonical form, minified, then converted back — a thin adapter layer, no optimization logic is duplicated. The Anthropic path is identity (zero work, byte-identical to before).

Pair-safety is preserved across the round trip: an assistant tool call plus its following role:"tool" replies become Anthropic tool_use + tool_result blocks, the pipeline drops such pairs together, and the reverse conversion never orphans a tool result from its call.

# proxy: point any of these at slimtoken; it detects the format from the path
export OPENAI_BASE_URL=http://127.0.0.1:8181/v1     # OpenAI clients → /v1/chat/completions
export OLLAMA_HOST=127.0.0.1:8181                    # Ollama clients → /api/chat
slimtoken serve --upstream http://127.0.0.1:11434   # → your local Ollama

# CLI: minify an OpenAI/Ollama body directly
slimtoken optimize -f openai  -i req.json
slimtoken optimize -f ollama  -i req.json

Local-model presets by VRAM

Recommended configs for common local models grouped by GPU VRAM tier, each with a usable context (KV cache + overhead eat into the nominal max). The reduction column is the live measured token drop the always-on pipeline achieves on the bloated payload — computed by the pipeline, not hand-waved (slimtoken presets --measure).

Reduction is config-dependent, not model-dependent — the pipeline rewrites the request regardless of which model consumes it, so every tier shows the same number (the always-on config on a bloated payload). On a typical session it's ~9%. Tune the config with the SLIMTOKEN_* env knobs, not by switching models.

Effective context window — dense vs MoE

Because slimtoken compresses input ~85%, a model's nominal context window holds far more raw conversation than its size suggests. The effective capacity is nominal_ctx / (1 − reduction). The presets below push each tier to the largest nominal context that fits fully in VRAM (q4_0 KV, flash attention, full GPU offload, --kv-unified) — computed by the backend optimizer, not asserted — and show the effective raw-token capacity with compression. Each tier has both a dense and a MoE/Mamba-hybrid option: hybrids (Qwen3.6-35B-A3B, LFM2.5-8B-A1B) have ~5 KB/token KV vs ~30 KB/token for dense, so they reach far larger contexts on the same VRAM.

slimtoken high-context                 # full table (all tiers, dense + MoE)
slimtoken high-context --vram-gb 16    # one tier
slimtoken high-context --vram-gb 16 --detail   # + the llama-server commands

The 16 GB MoE row is capped at 128 k — the proven-stable value on a 16 GB card (256 k OOMs at ub=2048; 128 k@ub512 measured 13.7 GB). The 8 GB MoE row is capped at 128 k too (256 k is a razor fit, ~+0.04 GB margin — any VRAM spike spills it; 128 k leaves ~0.67 GB headroom). The 4 GB MoE at 2-bit is a quality trade-off — the dense 3B is usually the better 4 GB pick. All configs use q4_0 KV (-ctk q4_0 -ctv q4_0) matching a proven local llama-server setup.

MCP server

slimtoken-mcp exposes the pipeline as MCP tools over stdio (the transport every local MCP client uses). It is a thin adapter: every tool imports and calls an existing core function — no optimization is reimplemented. The proxy and the MCP server are independent processes that share the same library.

optimize_messages, estimate_tokens, and inspect_budget accept a format field (anthropic / openai / ollama); non-anthropic bodies are normalized to canonical before the pipeline runs and returned in the caller's format.

Wire it into an MCP client's stdio config (example for Claude Desktop / claude_desktop_config.json):

{
  "mcpServers": {
    "slimtoken": {
      "command": "slimtoken-mcp"
    }
  }
}

The server speaks the MCP JSON-RPC 2.0 protocol (initialize → tools/list → tools/call), protocol version 2024-11-05, and is self-contained (stdlib only on top of slimtoken's existing deps). It does no optimization itself — every call dispatches to the core pipeline.

Agent Skill

The skills/slimtoken-optimizer/ directory is a packaged Agent Skill (static files a host agent runtime — Claude Code, ADK, Gemini CLI, Cursor — reads from disk on activation, not a running process). It is model-agnostic and works with local, cloud, or uncensored models: it rewrites the request, not the model.

skills/slimtoken-optimizer/
  SKILL.md                          # L1 description (~50 tok) + L2 body (<800 tok)
  references/optimization-policies.md  # full stage list + pair-safety rules (loaded on demand)
  scripts/optimize.py               # wrapper: CLI primary, MCP stdio fallback

The wrapper shells out to the slimtoken CLI when available, and falls back to a one-shot MCP stdio call (slimtoken-mcp) when only the MCP server is installed — so the skill works regardless of which surface the host has:

python3 skills/slimtoken-optimizer/scripts/optimize.py optimize -i req.json
python3 skills/slimtoken-optimizer/scripts/optimize.py presets --vram-gb 16 --measure
python3 skills/slimtoken-optimizer/scripts/optimize.py estimate -i req.json

Drop the skills/slimtoken-optimizer/ directory into your agent's skill search path and the host runtime surfaces it when a prompt matches "shrink / minimize / trim tokens / context too long".

Lossy stages

Two stages discard information. Tool-result compression is on by default (part of the always-on pipeline); output filtering is opt-in.

Config

Defaults are the recommended values. Set any to 0 to disable.

GET /metrics returns cumulative token counts + the t0–t4 latency buckets. TLS for cloud HTTPS upstreams is handled by httpx (SNI; optional mTLS via SLIMTOKEN_TLS_*; SLIMTOKEN_TLS_INSECURE=1 to skip verify). Lazy MCP — one stub tool per configured MCP server in ~/.slimtoken/lazy_mcp.json, the real server spawned on call — is available as a separate entrypoint.

Backend optimizer — the config-optimization stack

flowchart TB
    subgraph Fit[fit the model in VRAM]
        W[weights<br/>-ngl 999 full offload] --> K[KV cache<br/>-ctk/-ctv q4_0]
        K --> CB[compute buffer<br/>--kv-unified]
    end
    subgraph Speed[decode speed levers]
        FA[flash attention<br/>-fa on] --> UB[big ubatch/batch<br/>-ub N -b N]
    end
    C[context window<br/>-c N] --> Fit
    Speed --> Result[2-4× decode speedup<br/>~50-75% less wall-clock<br/>~2× context capacity]
    Fit --> Result

slimtoken config-optimizer [--model /path/to.gguf] [--vram-gb 16] [--model-size-gb 12.7]
                            [--kv-per-token 5120] [--native-ctx 262144]

config-optimizer inspects your GPU VRAM and model size, estimates weights VRAM, KV cache, and the compute buffer, then recommends llama-server arguments that fit without OOMing. It prints a ready-to-paste llama-server command plus CORTEXAGENT_* env exports. It changes nothing itself — recommend-only.

Total potential: versus a naive baseline (partial CPU offload + fp16 KV + no flash attention), enabling all of the above typically yields a 2–4× decode speedup (≈50–75% less wall-clock per token) and ~2× context capacity. These are typical llama.cpp ranges, not measurements taken by slimtoken — the real figure depends on your starting config. config-optimizer computes the largest safe values for your specific VRAM automatically.

⚠️ Estimate only — verify VRAM with nvidia-smi under a real prompt before trusting the margin. The compute buffer is calibrated for --kv-unified on a hybrid MoE; dense models or --kv-budget change the math.

Tests

python3 tests/test_all.py          # 161 checks — core pipeline + proxy + adapters + context presets
python3 tests/test_mcp_server.py   # 60 checks — MCP stdio server (all 8 tools)

Cover fence byte-identity, pair-safety, dedup, distill, ≥50% default reduction on a bloated payload, real-tokenizer counting (no whole-body serialize), single-pass equivalence, type-compressor pair-safety, output-filter truncation, async proxy end-to-end, /metrics latency buckets, fast-path byte-identical passthrough, and the full MCP stdio handshake + every tool + the error paths.

License

MIT, Copyright (c) 2026 greyok00. See LICENSE.