The Enabling Tools of AI Coding Agents

The article's premise is that more context does not mean better results: giving an agent thousands of files adds tokens, noise, cost, latency and chances to focus on the wrong code. The better question is how to give the agent exactly the context it needs. The tools below address that at different levels, from packing a repository into one file to letting the agent retrieve what it needs through search, symbols, graphs and MCP. Optimise for better context, not more.

Repomix packages a repository into one AI-friendly file with token counting, include and exclude controls, Git awareness and optional Tree-sitter compression. Gitingest turns a GitHub repository into text with a URL change, which is convenient for a quick look at an unfamiliar project. code2prompt builds a structured prompt from selected files, for task-specific context. The article advises against packing everything blindly: use include and ignore filters.

Aider's repository map lists important files, classes, functions and signatures and ranks them within a token budget (about 1,000 tokens by default, configurable with --map-tokens), so the model knows what exists and where. Semantic search such as grepai finds conceptually related code, not only exact string matches, and Sourcegraph provides code search, navigation and natural-language Deep Search across many repositories, branches and commits.

Serena gives coding agents semantic, symbol-level tools such as finding a symbol, finding its references and inserting after a symbol, exposed through MCP. Reading a whole 2,000-line file to inspect one method is wasteful; symbol retrieval lets the agent focus on the relevant symbol and its neighbours.

A code graph represents files, classes, functions and APIs as nodes and imports, calls, inheritance and references as edges, so an agent can ask what depends on a service and traverse the relationships instead of reading files. The article says graphs become more useful as repositories grow past thousands of files. The tools below index a repository into a persistent graph and serve it to agents, often through MCP.

DeepWiki generates documentation you can talk to for a repository, useful for onboarding. Greptile indexes a codebase and uses that understanding for AI code review. Riflet is a context builder that combines code with other sources, such as documents, before they reach the model.

Instead of a giant context file, MCP lets an agent call tools for search, symbols, references, file reads, Git history, tests and browsers, retrieving information when it needs it, much as a developer works. Persistent instruction files such as AGENTS.md and Cursor rules record architecture, commands and conventions once, so they are not repeated in every prompt.

The article's rule: do not ask the model to do what a deterministic tool does better. Run the type checker, the linter and git diff, git log, git blame and git grep yourself, then give the agent the facts to reason about, and feed back only the failures.

Neighbouring pages on this site.