The AI coding agent ecosystem now has 99 named tools, and counting.
Kevin Elliott’s AgentManager (agentmgr) ships with an embedded catalog of 99 agents, from Anthropic’s Claude Code and OpenAI’s Codex to obscure entries like abtop, oh-my-pi, and zerostack. The open-source CLI tool, written in Go, detects installed agents, checks for updates across npm, pip, Homebrew, and native installers, and manages them from a single terminal interface.
The number is the story. Ninety-nine agents is not a sign of a healthy, maturing market. It is a sign of fragmentation, duplication, and the early-stage chaos that precedes consolidation. Every major AI lab ships its own coding agent. Anthropic has Claude Code and Claude Squad. OpenAI has Codex. Google has Gemini CLI. GitHub has Copilot CLI. DeepSeek has DeepSeek CLI and DeepSeek-Reasonix. Alibaba’s Qwen team ships Qwen Code. Mistral has Mistral Vibe. Amazon has Amazon Q Developer CLI.
Then there are the dozens of independent projects: Aider, OpenCode, Cline CLI, Goose, Open Interpreter, Codebuff, Plandex, and 75 more. The AgentManager catalog lists each agent alongside its supported installation methods — some support three or four package managers, others ship only as a binary or a pip install.
AgentManager itself is a response to this sprawl. It does not build agents. It manages them. The tool detects what is already installed, checks for stale versions, and can install or update agents from the command line. A --refresh flag forces re-detection. A background helper runs as a system tray application with update notifications. The catalog is embedded into the binary and refreshable from a remote source.
The existence of this tool tells you something about the state of AI infrastructure. Package managers already exist. Homebrew, npm, pip, cargo, and Go modules all handle software distribution. But none of them track AI coding agents specifically, and none of them detect what is installed on a machine. AgentManager fills a gap that did not exist two years ago: the need for a meta-manager for AI development tools.
Kevin Elliott, the maintainer, built AgentManager to solve his own problem. The GitHub repository shows active development — the project uses Go 1.25, ships both a CLI and a TUI built with Bubble Tea, and exposes REST and gRPC APIs for integration. The catalog is open-source and accepts contributions via pull requests. Adding a new agent is a single entry in catalog.json.
The fragmentation is not accidental. Each lab wants its own agent to drive adoption of its underlying model. Anthropic pushes Claude Code to sell Claude API credits. OpenAI ships Codex to drive usage of GPT-5. Google bundles Gemini CLI to keep developers inside its ecosystem. The result is a dozen near-identical tools that all do roughly the same thing: take a natural language prompt and generate, edit, or debug code.
The duplication is costly for developers. A team using Claude Code for one project and Gemini CLI for another must track two separate update cycles, two different configuration formats, and two different CLI interfaces. AgentManager abstracts that away, but it is a bandage on a broken model.
The 99-agent catalog also reveals something about the economics of AI coding tools. Most of these agents are free to download. They are loss leaders for model access. Anthropic does not charge for Claude Code directly — it charges for the Claude API calls the agent makes. OpenAI does the same with Codex. The tool is free; the compute is not. This creates a strange incentive structure where labs compete on agent quality not to sell the agent, but to capture the inference spend.
Some independent agents, like Aider and OpenCode, are open-source and community-maintained. They do not have a model vendor to subsidize them. They survive on donations, grants, or by routing through the user’s own API keys. The economics are thin.
The fragmentation will not last. Markets this crowded consolidate. The question is which agents survive. The ones with the best models? The ones with the best user experience? The ones backed by the largest API vendor? Or the ones that integrate most seamlessly into existing workflows?
AgentManager points toward a possible answer: the winner may not be any single agent, but the layer that manages all of them. If the future is multi-agent — developers running Claude Code for prose, Codex for boilerplate, and Aider for refactoring — then the management layer becomes the moat. AgentManager is an early bet on that future.
The catalog is already 99 agents. It will be 200 by next year. The tooling to manage them will only become more important.