Slashing Token Costs in GitHub’s Agentic Workflows

GitHub’s Agentic Workflows quietly handle thousands of automated tasks across its repositories, keeping codebases clean and CI pipelines efficient. But behind the scenes, these workflows consume tokens at scale—sometimes without developers even realizing it. Since these automations run on fixed schedules with predictable inputs, optimizing their token efficiency is both feasible and impactful.

In April 2026, GitHub’s engineering team launched a systematic effort to audit and reduce token consumption across its own workflows. The initiative focused on three key areas: measuring usage accurately, identifying inefficiencies, and applying targeted optimizations. The results? Some workflows now use 40% fewer tokens without any change in functionality—a significant win for both cost and sustainability.

Tracking Every Token Spent

GitHub relies on hundreds of agentic workflows, each powered by different frameworks like Claude CLI, Copilot CLI, and Codex CLI. Historically, tracking token usage was fragmented. Each framework logged data in its own format, and historical records were often incomplete. This made it difficult to pinpoint where improvements were needed.

The breakthrough came from GitHub’s security architecture. Every agentic workflow runs through an API proxy that blocks direct access to authentication tokens. This proxy became the key to centralized logging. By capturing all API calls in a single normalized format, GitHub could now generate a token-usage.jsonl artifact for every workflow run. Each record includes:

• Input and output tokens
• Cache reads and writes
• Model and provider details
• Timestamps for precise tracking

With this data, engineers could reconstruct historical usage patterns, identify outliers, and prioritize optimizations based on real-world impact.

Automating the Optimization Process

Measuring token usage was only the first step. To sustain efficiency, GitHub built two automated workflows that analyze and refine their own operations:

Daily Token Usage Auditor This workflow aggregates token consumption data from recent runs, flagging anomalies like sudden spikes in usage or workflows that exceed expected turn counts. For example, a routine task that typically completes in four LLM turns might occasionally spike to 18—an immediate red flag for investigation.

Daily Token Optimizer When the Auditor flags a workflow, the Optimizer generates a GitHub issue with actionable recommendations. It scans the workflow’s source code and recent logs to pinpoint inefficiencies, such as redundant tool registrations or unnecessary API calls. Many of these issues would have gone unnoticed without automation.

Interestingly, these optimizers are themselves agentic workflows. Their token usage is tracked and reported daily, creating a virtuous cycle of self-improvement.

Trimming Unused Tools from MCP Servers

One of the most common inefficiencies uncovered was the inclusion of unused MCP (Model Context Protocol) tools. Agent runtimes traditionally bundle the full tool manifest with every request, even if only a fraction of the tools are ever called. For a GitHub MCP server with 40 tools, this can add 10–15 KB of unnecessary schema per LLM turn.

Workflow authors often start with a comprehensive toolset for convenience, but over time, most workflows stabilize around a narrow set of essential tools. The Optimizer cross-references tool manifests with actual usage logs to identify and prune unused tools. In test workflows, this reduced per-call context size by 8–12 KB, slashing thousands of tokens per run—without altering behavior.

Replacing MCP with GitHub CLI for Data Fetching

A more transformative optimization came from replacing MCP-based data retrieval with GitHub CLI (gh) commands. MCP tool calls require the agent to:

• Decide to invoke the tool
• Construct arguments
• Receive and process the response as part of the context

This entire process consumes tokens for schema, arguments, and responses—even though the data fetching itself could be deterministic and efficient.

GitHub implemented two strategies to eliminate this overhead:

Pre-agentic Data Downloads For predictable data needs (e.g., pull request diffs or changed files), workflows now run gh commands before the agent starts. Results are saved to workspace files, which the agent reads directly. This removes MCP tool-call overhead entirely and leverages the agent’s training in shell scripting for efficient processing.

In-Agent CLI Proxy Substitution For dynamic data fetches—where the agent determines what to retrieve at runtime—GitHub uses a lightweight HTTP proxy. This proxy routes CLI commands to GitHub’s REST API without exposing authentication tokens. The agent runs commands like gh pr view --json and receives structured data, just as a user would in a terminal. This approach reduces token usage while maintaining GitHub’s zero-secrets security model.

The Road Ahead: Sustainable Automation

GitHub’s efforts demonstrate that agentic workflows don’t have to be opaque or costly. By combining precise measurement, automated auditing, and targeted optimizations, teams can slash token usage without sacrificing functionality. The next phase will focus on expanding these techniques to third-party workflows and refining the optimizer’s recommendations.

As AI-driven automation becomes more pervasive, efficiency will be the difference between scalable systems and runaway costs. GitHub’s approach offers a blueprint for teams looking to balance innovation with sustainability in their workflows.