We cut your AI token costs by 70-95%. Measurably.
The Token Optimization Stack: 7 layers of compounding cost reduction. Model routing, prompt caching, session dedup, wire format encoding, context filtering, output compression, agent loop architecture. Most teams optimize one layer. We audit all seven.
Projects
Open source infrastructure for the agentic AI stack
Wire formats, code intelligence, MCP tooling, conformance testing. Everything production-grade, everything backed by data.
GCF (Graph Compact Format)
The AI-native wire format for structured data. 100% LLM comprehension on every frontier model, 90.7% on adversarial payloads. 53-71% fewer tokens than JSON. 1,700+ evaluations, 10+ models, 3 providers. Six implementations, zero runtime dependencies, MCP proxy, 43 billion+ lossless round-trips across 5 formats.
agent-lsp
Code intelligence infrastructure for AI agents. 65 tools, 30 CI-verified languages, 24 agent workflows. Single Go binary. Uses GCF as default output format.
mcp-assert
Conformance testing for MCP servers. 102 servers scanned, 34 bugs found, 12 upstream issues filed. ~28K downloads. Fuzz testing, schema linting.
knowing
Self-adapting code intelligence engine. 28 MCP tools, graph-native analysis, session deduplication, content-addressed identity layer. The system GCF was extracted from.
GCF Proxy
Wrap any MCP server with one command. JSON tool responses re-encoded as GCF before reaching the model. Zero code changes. Session stats track savings in real time.
GCP Emulator Suite
Local implementations of Google Cloud APIs. Secret Manager has 50K+ downloads, ranked #1 on Google/Bing/DuckDuckGo. KMS, IAM, Eventarc also available.
Consulting
The Token Optimization Stack
These layers stack multiplicatively
A $2.28 agent session becomes $0.36 after applying routing + caching + format encoding + effort control. Each layer addresses a different source of waste. Optimizing one does not diminish the opportunity in the others.Layer 1: Model Routing
Route simple tasks to cheap models (80% savings). RouteLLM, LiteLLM, or custom classifiers. Most teams use one model for everything.
Layer 2: Transport
Prompt caching (90% off repeated prefix), batch API (50% off async work), request coalescing. Zero code changes on Anthropic and OpenAI.
Layer 3: Session
GCF session deduplication: 92% savings by the 5th call. Delta encoding: 81% on re-queries. No other format can do this.
Layer 4: Format
GCF wire format: 53-71% fewer tokens on structured data. 100% LLM comprehension. Drop-in proxy or native encoding in 6 languages.
Layer 5: Context
RAG reranking, CLI output filtering (RTK: 60-90% savings), selective tool loading. Send less irrelevant data into the context window.
Layer 6: Output
Effort parameter (thinking token control), structured outputs, Caveman-style compression. Output tokens cost 3-6x more than input.
Layer 7: Architecture
Agent loop redesign, task budgets, parallel tool execution, progressive disclosure. The loop is where all optimizations compound most aggressively.
30 minutes. We'll show you your actual savings on your real data.
How it works
Step 1: Discovery Call (30 min, free)
Eight questions that reveal which of the 7 layers are costing you the most. We identify your batch-eligible workload, cache hit rate, routing opportunities, and structured data percentage.
Step 2: Token Audit (1-2 days)
Instrument one representative session. Measure total tokens per call, breakdown by category (system prompt vs history vs tool results vs output), cache hit rate, and structured data ratio. Deliver a report with projected savings per layer.
Step 3: Implement (1-2 weeks)
Deploy optimizations in priority order: prompt caching (day 1), format encoding (day 2), model routing (day 3-4), output compression (day 5), agent loop redesign (week 2). Measured before/after on every change.
Permanent cost reduction
Upstream
27 merged PRs across the ecosystem
#6 contributor to mcp-go (8.7K stars). Data corruption fixes, panic recovery, SDK hardening, spec compliance, transport bugs.
Anthropic
MCP Go, Python, PHP SDKs + servers (85K+ stars)
go-containerregistry: OCI artifact corruption fix (3.8K stars)
GitHub
github-mcp-server (16K stars)
Grafana
3 PRs merged: feature + panic fix + error propagation (74K stars)
etcd (CNCF)
gRPC error code fix (51K stars)
mark3labs (mcp-go)
9 PRs merged, #6 all-time contributor (8.7K stars)
Spending $5K+/month on LLM APIs?
Most teams are paying 3-5x more than they need to. We audit all 7 layers, show you the math, and implement the fixes. Typical result: 70-95% cost reduction in 2 weeks.