Skip to main content
Coding agents write code. Coda handles the work around it: PR reviews, issue triage, feature planning, debugging. It lives in Slack as a teammate and works against your codebase on disk.

​
The system

Coda is a multi-agent system:
MemberRoleTools
ExplorerReads code, traces call chainsCodingTools (read-only), GitTools, GithubTools, ReasoningTools
CoderWrites code in isolated worktrees, opens PRsCodingTools (full), GitTools (worktree), GithubTools, ReasoningTools
PlannerBreaks features into ordered GitHub issuesCodingTools (read-only), GitTools, GithubTools (issues), ReasoningTools
ResearcherHits the web for framework docs, library APIs, CVEsParallelTools, ReasoningTools
TriagerLabels, comments, closes issues based on the actual codeCodingTools (read-only), GithubTools (issues), ReasoningTools
LeaderRoutes the request, posts to SlackSlackTools
The Coder works in a coda/<branch> git worktree off main. A human reviews and merges. Coda never merges. Worktrees clean up on restart.

​
How Coda runs

Code is searched on disk. Grep, find, read against actual files, synced every five minutes. No vector indexing. No stale embeddings. Three scheduled tasks ship by default:
TaskCadenceWhat it does
Repo SyncEvery 5 minutesPulls latest changes from every configured repo
Daily DigestEach morningPosts what merged yesterday, PRs waiting, issues opened, what’s gone stale
Issue TriageDailyReviews open issues against the actual code, labels them, posts a summary
Coda learns your conventions through Agno’s Learning Machine. Week 1 you get generic patterns. Week 4 it follows your service layer style, error handling, naming, and logging.

​
Next

Setup → takes about five minutes and points at the agno repo by default.