v0.1.0 — Beta · BYOK · 6 providers

The AI coding agent
for production codebases.

DevPilot reads repositories, searches code semantically, applies precise patches, runs tests, and opens pull requests — autonomously. Bring your own API keys. Connect any provider.

avg latency
1.1s
tokens/run
820
patch accuracy
98.4%
providers
6
tools
6
devpilot — agent
Architecture

Full-stack agent infrastructure, not a chatbot wrapper.

DevPilot runs an autonomous agent loop with real tool access. The system retrieves precise code context using semantic embeddings (RAG), reads and writes actual files via GitHub API, executes shell commands, and routes LLM calls across providers based on task type.

Agent loop with configurable max steps
ChromaDB-backed semantic code search
Unified diff generation + GitHub write-back
Provider routing: task-type aware
Supabase auth + per-user BYOK key storage
Developer
DevPilot UI
Agent Loop
RAG / Embeddings
GitHub API
LLM Providers
Core
UI / Auth
LLM Providers
Capabilities

Six tool primitives.
Infinite task surface.

Every agent run is a composition of these six tools. The agent decides which to call, in what order, based on the task. Tool calls are logged in real time in the execution panel.

search_repo()

Semantic code search via embeddings

read_file()

Read file contents from any branch

write_file()

Write modified file back to repo

list_files()

Enumerate repository structure

run_command()

Execute shell commands and tests

generate_patch()

LLM-powered unified diff generation

Agent Workflow

Deterministic by default.
Adaptive when needed.

The agent loop is configured with a maximum step count. Each step has a defined action and produces a logged observation. You see exactly what the agent does — no black boxes.

01

Connect Repository

Point DevPilot at any GitHub repository. The agent indexes code semantically using embeddings for precise context retrieval.

github.com/org/repo@main
02

Describe the Task

Write a natural-language instruction. Refactor, fix, document, test — any coding intent expressed in plain language.

"Add rate limiting to /api/auth endpoints"
03

Agent Executes

The agent loop runs autonomously: search → read → reason → patch. Each step is logged and inspectable in real time.

4 steps · 1.4s · 891 tokens
04

Review & Apply

Inspect the unified diff before committing. Accept, reject, or iterate. The agent can open PRs directly from the interface.

feat/rate-limiting → open PR
Bring Your Own Keys

Your keys. Your models.
Your cost control.

DevPilot stores your API keys encrypted per-user in Supabase. When you run an agent task, your key is retrieved and sent directly to the provider — DevPilot never proxies or logs your requests. Platform fallback keys are used only when no user key is found.

# Key resolution order
1. user.api_keys[provider] → BYOK
2. env.PLATFORM_KEY → fallback
3. raise HTTPException(400) → no key
OpenAI
gpt-4ogpt-4-turbo+1
Anthropic
claude-3.5-sonnetclaude-3-haiku
Groq
llama-3.1-70bmixtral-8x7b
Gemini
gemini-1.5-progemini-flash
DeepSeek
deepseek-coderdeepseek-chat
OpenRouter
any model via proxy
Security

Zero-proxy key handling

API keys are retrieved server-side only at runtime. Never stored in logs or sent to client.

JWT auth via Supabase

All API routes require a valid Supabase JWT. User identity is verified on every request.

Scoped GitHub access

Repository access uses OAuth tokens with minimum required scopes. Read by default, write only when explicitly applied.

Ship faster with an agent that
actually understands your code.

Connect a repository, describe a task, and watch DevPilot execute — with full observability at every step.