Token Savings

Token Savings (Measured)

AgentCTX’s token savings are measured, not estimated. These benchmarks come from real agent sessions comparing raw MCP tool usage against AgentCTX-managed sessions.

Headline Numbers

Category	Without AgentCTX	With AgentCTX	Savings
Tool descriptions (70 tools)	9,950 tokens	850 tokens	91%
Tool invocations (150 calls)	7,350 tokens	1,800 tokens	76%
File search + read	16,925 tokens	1,290 tokens	92%
Documentation search	11,900 tokens	890 tokens	93%
Total per session	46,125 tokens	4,830 tokens	90%

How Each Saving Works

Tool Description Compression (91%)

Without AgentCTX, every tool’s full JSON schema is injected into the context window:

70 tools × ~142 tokens per schema = 9,950 tokens

With AgentCTX, agents discover tools via ?t queries and only load schemas they need:

Gateway tool (~300 tokens) + on-demand inspection (~550 tokens) = ~850 tokens

Tool Invocation Efficiency (76%)

Without AgentCTX, each tool call carries the full MCP JSON-RPC envelope:

Request + response × 150 calls = 7,350 tokens

With AgentCTX, CTX statements are compact:

>t github.issues.create title="Fix SSE" = ~25 tokens vs ~49 tokens for raw MCP

Static Content Deduplication (92–93%)

The content-addressed store (CAS) deduplicates static content:

File contents stored once by SHA-256 hash
Subsequent references use 3-token hash pointers instead of full content
Documentation chunks indexed and retrieved by relevance, not dumped wholesale

Methodology

See Methodology for full details on:

Test environment and models
Workload definitions
Measurement instrumentation
Statistical validity