Your coding agent is burning through GitHub API tokens — here's the fix
Your coding agent just hit GitHub's rate limit again. It's the third time this week, and you're starting to wonder if letting an AI loose on your repos was a mistake.
Here's what's happening: your agent is treating GitHub like Google. Every time it needs to understand a codebase, it's making dozens of API calls to fetch file contents, check commit history, and scan issues. A simple "add logging to the user service" task turns into 200+ API requests.
The fix isn't smarter caching or a bigger API budget. It's teaching your agent to work like a developer who pays for their own tokens.
The Local-First Pattern
Instead of fetching files through the API, your agent should clone repos locally and work with the filesystem directly. Here's the pattern that cut my GitHub API usage by 90%:
// Bad: API-heavy approach
await github.repos.getContent({owner, repo, path: 'src/user.js'})
await github.repos.getContent({owner, repo, path: 'src/utils.js'})
await github.repos.getContent({owner, repo, path: 'package.json'})
// ... 47 more API calls
// Good: Local-first approach
exec(`git clone https://github.com/${owner}/${repo} /tmp/workspace`)
const userFile = fs.readFileSync('/tmp/workspace/src/user.js', 'utf8')
const utilsFile = fs.readFileSync('/tmp/workspace/src/utils.js', 'utf8')Your agent gets the same information, but file reads are free and instant.
When to Use the API (Sparingly)
Reserve GitHub API calls for actions that require them:
- Creating PRs and issues — You can't do this locally
- Reading recent activity — Comments, reviews, recent commits
- Checking CI status — Build results, deployment status
- Managing branches — Creating, merging, deleting branches
Everything else should happen locally. File contents, directory structure, git history, searching code — all of this is faster and cheaper when your agent works with a local clone.
Security note: If your agent is working with private repos, make sure your workspace cleanup is bulletproof. Set up auto-deletion after each session and never persist sensitive repos to disk.
The Workspace Pattern
Give your coding agent a dedicated workspace directory that it manages:
const workspace = '/tmp/agent-workspace'
// Start each coding session fresh
fs.rmSync(workspace, {recursive: true, force: true})
fs.mkdirSync(workspace, {recursive: true})
// Clone what you need
exec(`git clone ${repoUrl} ${workspace}/${repoName}`)
// Work locally
process.chdir(`${workspace}/${repoName}`)
// Push changes when ready
exec('git add . && git commit -m "Agent: Add logging" && git push')This pattern works especially well with tmux-based coding sessions where your agent can maintain context across multiple terminal windows while working entirely with local files.
The result? Your agent becomes faster (no API latency), cheaper (no token costs for reading files), and more reliable (no rate limits blocking progress). Plus, it can work with private repos without burning through your GitHub Enterprise API quota.
If you're running coding sessions that keep hitting API limits, this workspace approach will transform how your agent operates. It's the difference between an agent that costs $50/month in API calls and one that costs $5.