GPTCode

AI coding agents should produce evidence, not just answers.

GPTCode is an open-source coding CLI for investigating, planning, implementing, reviewing, and verifying software with multiple AI models. It keeps repository knowledge, engineering constraints, and executable verification close to the code.

Website · Architecture · Documentation · Engineering thesis · Evaluation essay

The thesis

Language models generate possibilities. Repositories define constraints. Verification establishes truth.

GPTCode separates those responsibilities into explicit, inspectable stages:

Repository knowledge
  Skills · contracts · tests · documentation
                    │
                    ▼
Research → Plan → Implement → Review → Verify
                    │
                    ▼
OpenAI · Gemini · Groq · OpenRouter · Ollama

The workflow is the source of truth. Models are interchangeable execution engines.

Install

GPTCode requires Go 1.24 or newer:

go install github.com/jadercorrea/gptcode/cmd/gptcode@latest
gptcode setup

Connect a provider:

gptcode key openrouter

Provider credentials remain local to your GPTCode configuration. Review the commands and provider policies before using the CLI with confidential code.

A real workflow

gptcode research \
  "How does session expiration work, and is concurrent access safe?"

gptcode review session/store.go \
  --focus "concurrency correctness and public API stability"

gptcode do \
  "Fix the session store concurrency bug without changing its public API. Verify with go test -race ./..."

The published demonstration uses a real Go fixture and ends with executable evidence rather than a model assertion:

Research   → identifies the shared-state boundary
Review     → confirms unsynchronized map access
Implement  → preserves the public API and adds synchronization
Verify     → go test -race ./...

See the auditable fixture and the recording source.

Verify the evidence

Run the same deterministic checks used in CI:

make evidence

This executes the public fixture with Go's race detector and fails unless it retains 100% statement coverage:

ok  gptcode/examples/sessionstore  coverage: 100.0% of statements
public examples verified: race detector passed; statement coverage 100.0%

To run the complete repository quality gate:

make verify

That gate builds the CLI, runs go vet, executes the short test suite, checks the public project contract, and runs the race/coverage evidence.

The complete claim-to-evidence matrix is documented in QUALITY.md.

Repository-native skills

Skills are version-controlled engineering instructions loaded from the repository. They can define language conventions, architectural boundaries, testing practices, and review criteria:

gptcode skills list
gptcode skills show go
gptcode skills install ruby

Core commands

Command	Purpose
gptcode research "question"	Investigate code and produce repository-grounded findings
gptcode plan "task"	Turn ambiguity into an inspectable implementation plan
gptcode do "task"	Implement a bounded change with tool-backed verification
gptcode review [path]	Review correctness, security, and contract stability
gptcode chat	Explore a repository interactively
gptcode skills list	Inspect repository-native engineering guidance

Use gptcode --help for the complete command surface.

Design principles

• Repository-centered: project constraints outrank temporary prompt context.
• Explicit stages: research, planning, implementation, review, and verification remain inspectable.
• Provider-independent: route work according to capability, latency, cost, and privacy requirements.
• Executable evidence: tests, linters, race detection, and project commands decide whether a change is acceptable.
• Honest boundaries: model output is a proposal, never proof of correctness.

Limitations

GPTCode is an independent research-driven project, not a hosted service or a guarantee that model-generated changes are correct.

• Model-backed workflows require credentials for a supported provider.
• Results vary by model and repository context.
• The 100% coverage claim applies to the published sessionstore fixture, not to the entire legacy codebase.
• Some experimental command surfaces predate the current repository-centered architecture and are being evaluated or retired.
• Always review changes and run the repository's own verification commands.

Development

git clone https://github.com/jadercorrea/gptcode.git
cd gptcode
make verify

Releases are intentional: pushing a version tag runs the full verification gate before GoReleaser publishes checksums and platform archives to GitHub.

Read CONTRIBUTING.md, SUPPORT.md, and SECURITY.md before opening a contribution or reporting a vulnerability.

Project status

GPTCode is actively maintained as an open-source engineering laboratory for reliable AI-assisted software development. The roadmap is public in _roadmap.md.

Created by Jader Correa, a principal engineer building AI agents, developer tools, and distributed systems.

License

MIT