AgentOne Token Compression

AgentOne Token Compression

by Iterate.ai

Cut the tokens Claude ingests from tool output — ~75% overall on a mixed Claude Code workload (big on JSON, Bash, Grep and repetitive logs; little on prose or varied logs). Dropped code bodies recoverable via on-demand retrieve. Drop-in. Zero-config. Runs locally.

Works in: Claude Code · Claude CLI · Claude Desktop
Part of the AgentOne platform.

---

What it does

Every time Claude reads a file, runs Bash, executes Grep, fetches a URL, or you paste a chunk of content, those tokens enter Claude's context and get billed on every subsequent turn. AgentOne Token Compression intercepts content at the point of insertion and compresses it so the bill — and the context pressure — stays low. Savings are highly content-dependent: large on verbose or repetitive tool output (JSON, Bash/Grep results, near-identical log lines) and small on prose or logs where every line differs. On the mixed benchmark workload below the overall reduction is ~75%. Dropped code bodies are recoverable via on-demand retrieve.

Measured with the plugin's benchmark harness (ships in the development source repo). See the Benchmarks section for the raw output.

┌─────────────────────────────────────────────────────────────────┐
│  Read on big JSON file        8.9 kB  →   1.3 k tokens  ⚡ 40%   │
│  Bash output (npm install)    5.9 kB  →     161 tokens  ⚡ 89%   │
│  Grep on codebase             4.1 kB  →     152 tokens  ⚡ 88%   │
│  WebFetch HTML page           2.2 kB  →     364 tokens  ⚡ 47%   │
│  Long log file *              6.8 kB  →       0 tokens  ⚡ 100%  │
│                                                                 │
│  Overall:  8.0 k tokens  →  2.0 k tokens   ⚡ 75% saved          │
│  Hook latency: ~66 ms                                           │
└─────────────────────────────────────────────────────────────────┘

* Best case: near-identical log lines. Logs with unique per-line values (timestamps, request ids) compress far less — sometimes ~0%.

It also caches every tool result and prompt, so a repeat read of the same file (or a near-identical prompt) returns the same compressed form quickly without recomputing it — a cache hit saves the recompute work (~7–8% additional token savings on re-reads), not 100% of the tokens.

---

Three ways to use it

Surface	Mode	Auto-compresses?
Claude Code (TUI)	Hooks via ~/.claude/settings.json	✅ Yes — every tool call
Claude CLI	Same hooks as Claude Code	✅ Yes — every tool call
Claude Desktop	MCP server	On-demand (Claude calls the tools)

---

Quick install

Claude Code — plugin marketplace (recommended)

/plugin marketplace add IterateAI/compression
/plugin install agentone-token-compression@iterate-ai

Restart Claude Code. Hooks auto-compress every tool call — no further setup.

Claude Code & CLI (npm — alternative)

npm install -g @iterate.ai/agentone-token-compression
agentone-tc install

Restart Claude Code. Done.

Claude Desktop (MCP server via npx — on-demand)

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "agentone-token-compression": {
      "command": "npx",
      "args": ["-y", "@iterate.ai/agentone-token-compression", "mcp"]
    }
  }
}

npx -y fetches the package on first run; no global install needed. Fully quit Claude Desktop and relaunch.