context-mode

Overview

context-mode targets the dirty secret of long agent loops: most of the context budget is spent re-presenting the output of tool calls — file reads that produced 800 lines, shell commands that spilled stack traces, search results with 40 hits when the agent only acted on one. context-mode sandboxes that output, surfaces only the relevant slice to the LLM, and lets the rest stay out of context.

The empirical claim is 60-98% context-spend reduction on real coding workflows, with no measurable degradation in agent decision quality. That’s a large enough delta to be worth a serious test on any team paying agent-loop bills at scale.

Why It Matters

By Q2 2026, context optimization has emerged as a distinct runtime category, separate from the model layer and the harness layer. context-mode and rtk-ai/rtk trended in the same week with overlapping pitches, which is the GitHub-trending pattern that usually precedes a category consolidation.

The economic case stacks with the DeepClaude shim: if DeepClaude cuts the per-token cost ~17×, and context-mode cuts the token count 60-98%, the combined effect is a 50-1000× reduction in agent-loop spend versus stock Claude Code on Anthropic API. That’s a real magnitude shift, not a marginal optimization.

How It Works

• Tool-call interception. context-mode proxies the agent harness’s tool-call layer. When the agent invokes read_file, bash, grep, or similar, the full output goes to context-mode’s sandbox.
• Selective surfacing. A second-pass prompt or rule-based selector identifies the relevant slice (lines containing the diff target, the error from the stack trace, the matching grep hits). Only that slice gets returned to the agent’s conversation.
• Audit trail. The full output is logged to disk with a content-addressable hash. The agent can request the full version on demand if its first pass missed.

Use Cases

• Heavy agent loops on Anthropic API where token spend dominates the bill.
• Multi-file refactors that read tens of files; usually only a handful of lines per file matter.
• Shell-driven workflows where command output is verbose (build logs, test output, package-manager noise).
• Cost-sensitive teams running on DeepClaude or other discount endpoints, stacking the savings.

Pairs With

• DeepClaude — multiplicative cost reduction.
• rtk — competing project on the same thesis; benchmark both.
• jcode, Claude Code, Cursor, Codex — the harnesses context-mode plugs into.

Verdict

If you’re spending more than $200/month on Claude API tokens for agent loops, run context-mode for a week and measure. The author’s claim is large enough that even a fractional improvement justifies the install cost; the rtk peer-project’s overlapping numbers suggest the underlying technique is real. Expect this category to consolidate into one or two winners by Q3 2026 — context-mode is one of the two front-runners.