If you love VS Code but wish your AI coding experience was more structured and accurate, this tutorial is for you.
I built BitFrog Copilot — a free, open-source VS Code extension that adds 7+1 specialized AI agents to GitHub Copilot Chat. Instead of one general-purpose agent doing everything, each agent has a clear role: brainstorming, planning, TDD execution, debugging, code review, mentoring, and UI design.
In this post, I'll show you how to install it and walk through three real scenarios so you can see it in action.
Install in 30 Seconds
Option 1 — VS Code Marketplace:
Open VS Code
Go to Extensions (Ctrl+Shift+X)
Search "BitFrog Copilot"
Click Install
Option 2 — Agent Plugin:
Extensions sidebar → Agent Plugins → Search "bitfrog-copilot"
That's it. No API keys, no config files.
Requirements: VS Code with GitHub Copilot enabled. Works with GPT-5.4, Claude Opus 4.6, Gemini 3.1 Pro — your choice.
Meet the Agents
AgentWhat it does@bitfrogMain router — figures out what you need and sends you to the right agent@bitfrog-brainstormExplores ideas and designs before you write code@bitfrog-planMaps dependencies, decomposes tasks into ordered steps@bitfrog-executeTDD implementation — writes tests first, then code@bitfrog-debugDiagnoses root causes using a structured four-method approach@bitfrog-reviewThree-level code review: spec → quality → intent@bitfrog-mentorTeaches through questions — never gives direct answers@bitfrog-ui-designUX research and interface design
Scenario 1: Build a Feature with TDD
Let's say you need to add a password validation function.
Step 1 — Brainstorm the requirements:
@bitfrog-brainstorm I need a password validation function.
It should check length, special characters, and common passwords.
The agent won't just say "sure, here's a function." It will ask you questions: What's the minimum length? Do you need Unicode support? Should it check against a breach database? It investigates the real problem before proposing solutions.
Step 2 — Plan the implementation:
@bitfrog-plan Let's implement the password validator we just designed.
The plan agent maps out dependencies and gives you an ordered task list. Each task is a handoff — you decide when to move forward.
Step 3 — Execute with TDD:
@bitfrog-execute Start with task 1 from the plan.
The execute agent writes tests first, then implementation. Not because it's forced to — because it understands why TDD matters. It spawns parallel sub-agents for independent tasks when possible.
Scenario 2: Debug a Tricky Bug
You've got a bug where API responses are randomly empty. Instead of just reading the error message and guessing:
@bitfrog-debug Our /api/users endpoint sometimes returns empty
arrays even though the database has data.
The debug agent uses a four-diagnostic-method approach (borrowed from traditional Chinese medicine):
Observe — What are the visible symptoms?
Listen — What do the logs and metrics say?
Ask — What changed recently? What's the pattern?
Examine — Check the actual code path and data flow
It diagnoses the level of the problem (is this a network issue? a race condition? a query bug?) before suggesting a fix. No more guessing.
Scenario 3: Code Review Before Merging
You've finished a feature and want a thorough review:
@bitfrog-review Review my changes in the auth module.
The review agent runs three passes:
Spec compliance — Does it do what it's supposed to do?
Code quality — Are there maintainability issues, performance problems, or security concerns?
User intent — Does it actually solve the user's problem, or just implement the ticket?
That third pass is where most reviews fall short. The agent asks: "The spec says add rate limiting, but is 100 requests/minute the right limit for this use case?"
The Workflow: Handoffs, Not Chaos
The agents aren't isolated. They form a workflow:
brainstorm → plan → execute → review
↕
debug
Each transition is a handoff button — the current agent suggests the next step, you decide when to proceed. debug and mentor can be called independently at any time.
This is the key difference from vanilla Copilot: instead of one agent context-switching between planning, coding, and reviewing, each agent stays focused on what it's best at.
Why Chinese Philosophy?
Most AI coding agents use hard rules: "you MUST write tests," "you MUST NOT skip root cause analysis." This works, but the model follows the letter and misses the spirit.
BitFrog uses five philosophical principles to shape how agents think:
格物致知 (Investigate First) — Understand the real problem before acting
知行合一 (Unity of Knowing and Doing) — If you know you should write tests, write them. No excuses.
辨证论治 (Diagnose Before Treating) — Same symptom, different root causes. Diagnose first.
中庸之道 (Right Measure) — Every action has a right amount. Don't over-engineer, don't under-engineer.
三省吾身 (Three Reflections) — Quality comes from reflecting on your process, not checking boxes.
Sounds weird? Yeah. Works surprisingly well.
Try It Now
GitHub: github.com/rainyulei/bitfrog-copilot
VS Code Marketplace: BitFrog Copilot
License: MIT — fully open source
Star the repo if you find it useful. Issues and PRs welcome.
Top comments (0)