# Token savings

URL: https://agentvera.dev/en/features/token-savings/ · Plan: Free

The base context is re-read on every turn, so a small saving compounds over long conversations. AgentVera lets you choose a context profile per agent, shows the context growing in the pane and compacts at your threshold.

**Context profiles.** Measurements are the context sent on the first request of a new session (Claude Code 2.1, Codex 0.155).

- Standard: Claude ≈34.4k, Codex ≈13.3k tokens.
- Balanced: only the Bash, Read, Edit, Write, Glob and Grep tools. Claude ≈23.7k tokens (−31%).
- Lean: user plugins, hooks, skills and external MCP are off too. Claude ≈13.3k (−61%), Codex ≈9.7k tokens (−27%).

**Auto-compact for every agent.** Measurements showed that most of the token cost is context re-read on every turn. So auto-compact is on by default: when an agent’s context passes the threshold (150k tokens by default), the conversation is summarised with the CLI’s own command (/compact, /compress, /summarize, /smol). The pane shows live context size, and you can compact by hand too.

- Context size is measured from each CLI’s own history, so the context chip and auto-compact work for agents beyond Claude Code and Codex.
- Claude Code, Codex, Hermes and Droid also compact inside long turns.
- A message to an agent idle for over an hour first compacts a large conversation, whose prompt cache has expired anyway.
- A read guard asks for an outline of big files first and refuses, once, a re-read of lines of an unchanged file it has already shown; new agents are told to read economically.
- Prompt revision uses a cheaper model, and the Transform node in flows passes on only the part of a long reply you need.

**Compress command output (RTK).** AgentVera ships the open-source RTK (Apache-2.0). Output from agents’ test, build, git and similar commands is shortened before it reaches the model, typically 60–90% fewer tokens. The RTK switch under Settings → Token savings is on by default, and changes apply to agents started afterwards.

- In Claude Code, opencode, Qwen Code, Gemini CLI and Copilot CLI, commands are rewritten through a hook: git status → rtk git status. Commands RTK doesn’t know are left alone.
- Codex and the other local agents get rtk on their PATH and are told to prefix commands with it, using rtk proxy when they need raw output.
- RTK telemetry is always off.
- Savings show under Settings → Token savings for this computer, and in the AI ROI tab per project.

**Usage limits and statistics.** Settings → Usage limits shows what’s left for each Claude and Codex account, window by window: 5-hour, weekly and, for Claude, Sonnet and Opus weekly, each with its reset time. The status bar refreshes these about once a minute. The Statistics tab in the project’s Details looks back instead.

- Work time, tokens, prompts and sessions for today, this week, this month, this year or all time.
- Claude and Codex token usage, with input, output, cache read and cache write shown separately, and your most used models.
- A day-by-hour heatmap and the 20 latest sessions in the chosen range.
- The data is read from local Claude Code and Codex transcripts on this computer.

**Models, tools and the map.** Automatic model rates each task simple, normal or hard and gives small jobs to a cheaper model. Smart tools cut Claude Code’s repeat file reads, full test runs and repeated commands. The project map lets agents find the relevant code without reading file after file.

- **Where do I see my remaining limits?** Settings → Usage limits shows, for each Claude and Codex account, what’s left in the 5-hour and weekly windows and when each resets.
- **What does the Lean profile turn off?** It narrows the tool set and doesn’t load user-level plugins, hooks, skills or external MCP. AgentVera’s own MCP servers are still added.
- **Is there a Balanced profile for Codex?** No. Codex has Standard and Lean.
- **Are these numbers the same in every project?** No. They were measured on the first request of a new session; your project’s own instruction files and tools change the result.
- **Which agents does RTK work with?** Every local agent: through hooks in Claude Code, opencode, Qwen Code, Gemini CLI and Copilot CLI, through instructions in the others. RTK ships with the macOS, Linux and Windows (x64) builds. Settings → Token savings says which agents each feature covers.
- **Is auto-compact on by default?** Yes. Unless you chose otherwise, it is on with a 150k-token threshold and works in every CLI that has a compact command. Change the threshold under Settings → Token savings.

---

**About AgentVera.** AgentVera is a desktop app for macOS, Windows and Linux, built by Solviera Teknoloji. It runs 18 AI coding agents (Claude Code, Codex, Gemini CLI, Copilot, Cursor, Antigravity, opencode, Hermes, Qwen Code, Kimi Code, Kiro, Amp, Droid, Goose, Vibe, Crush, Aider, Cline) side by side, per project, in real terminals with the user's own CLI installs and logins. The free plan runs 2 agents at once. Download: https://agentvera.dev/en/download/ · Pricing: https://agentvera.dev/en/pricing/
