github-agent

Ask about your GitHub repos, pull requests, and issues in plain English. Pick which AI model does the thinking — OpenAI, Claude, or DeepSeek.

your prompt (CLI or browser chat)
               │
               ▼
┌─────────────────────────────┐
│        github-agent         │   reads your prompt
└──────────────┬──────────────┘
               │
               ▼
┌─────────────────────────────┐
│           bridge            │   openai / claude / deepseek — talks to the AI model
└──────────────┬──────────────┘
               │
               ▼
┌─────────────────────────────┐
│ GitHub's remote MCP server  │   reads/writes your repos
└──────────────┬──────────────┘
               │
               ▼
an answer, or a posted PR comment

What you need

• Node.js 18+
• A GitHub personal access token
• An API key for at least one of: OpenAI, Anthropic (Claude), DeepSeek

Use it: command line

GITHUB_TOKEN=<your-token> ANTHROPIC_API_KEY=<your-key> \
  npx @clickonsearch/github-agent --provider claude "list my 5 most recently updated open PRs"

Swap in OPENAI_API_KEY + --provider openai, or DEEPSEEK_API_KEY + --provider deepseek, to use a different model. --provider can be left out if you set PROVIDER instead.

Prefer not to put keys on the command line? Put the same variables in a .env file in whichever folder you run the command from — it's picked up automatically.

Use it: browser chat

GITHUB_TOKEN=<your-token> ANTHROPIC_API_KEY=<your-key> \
  npx -p @clickonsearch/github-agent github-agent-serve

Then open http://localhost:3000 in your browser and start chatting — pick your provider from the dropdown at the top.

Use it: embedded in another tool

The browser chat server speaks the AG-UI protocol (POST /agent, streaming SSE) — any AG-UI-compatible client can drive it as a backend, not just the bundled page. Point your own frontend at http://localhost:3000/agent once github-agent-serve is running.

Skills

Add --skill code-review (or any other skill from agent-skills) to change how the agent reasons, without changing what it can do:

GITHUB_TOKEN=<your-token> ANTHROPIC_API_KEY=<your-key> \
  npx @clickonsearch/github-agent --skill code-review "review pull request #12 in owner/repo and post your findings as a comment on it"

Point it at a real pull request (owner/repo + PR number) if you want the findings posted back to GitHub — the agent can only comment on something it can look up. A bare local diff has nothing to comment on, so the agent just returns the findings as text instead.

Config reference

| Env var | Required | Default | | -------------------- | ----------------------- | -------------------------------------- | | GITHUB_TOKEN | yes | | | PROVIDER | no | openai | | OPENAI_API_KEY | if using OpenAI | | | OPENAI_MODEL | no | gpt-4.1 | | ANTHROPIC_API_KEY | if using Claude | | | ANTHROPIC_MODEL | no | claude-sonnet-5 | | DEEPSEEK_API_KEY | if using DeepSeek | | | DEEPSEEK_MODEL | no | deepseek-chat | | GITHUB_MCP_URL | no | https://api.githubcopilot.com/mcp/ | | GITHUB_MCP_TOOLSETS | no | all | | PORT | no, browser chat only | 3000 |

For developers

Running from source (working on this agent's own code, not just using it):

# from the repo root, once
npm install
npm run build

# then, in this folder
cd agents/github-agent
cp .env.example .env    # fill in GITHUB_TOKEN + a provider key
npm run dev -- --provider claude "list my open PRs"
npm run serve            # browser chat, same as github-agent-serve above

This agent talks to GitHub's remote MCP server through whichever provider bridge you pick (openai / claude / deepseek) — all three expose the same interface, so swapping providers doesn't change how GitHub access works.

GitHub's remote MCP server groups its tools into toolsets (repos, issues, pull_requests, actions, ...) and only exposes an undocumented default subset unless asked for more, via the X-MCP-Toolsets header. This agent requests all by default so comment/review tools are always available; narrow it with GITHUB_MCP_TOOLSETS (comma-separated) if you want to restrict what the model can do.

Multi-turn memory in the browser chat is handled by flattening prior messages into the prompt (the underlying provider calls are single-turn).

Adding a new provider: add a runWith* function in src/providers.ts, add its name to PROVIDERS — CLI flags, env var checks, and the chat UI dropdown all read off that list automatically — then document its env vars above.

Using it as a library:

import { runGithubAgent } from "./src/index.js";

const answer = await runGithubAgent("claude", {
  prompt: "list my open PRs",
  githubToken: process.env.GITHUB_TOKEN!,
  githubMcpUrl: "https://api.githubcopilot.com/mcp/",
});