gdmcode 0.3.4

gdm: AI coding agent for professional developers

gdm — Open-Source AI Coding Agent

A terminal-native coding agent that reads your codebase, reasons over it, and edits files autonomously — with configurable autonomy, local model support, and a full audit trail.

---

Quick start

pipx install gdmcode
gdm version
gdm login grok        # or: gemini | codex
cd my-project && gdm "fix the login bug"

If you cannot reach a model provider directly, start gdm and run /proxy, then paste the relay token with /proxy token. The provider API key stays encrypted locally; the relay should not persist provider keys.

See docs/quick-start.md for a step-by-step walkthrough.

---

Why gdm?

• Local-model support — run fully air-gapped with Ollama or vLLM; no data leaves your machine.
• Autonomy slider — five levels from "ask everything" to "fully autonomous"; start low, build trust.
• Open source and auditable — MIT licensed, hash-chained audit log, team policy enforcement.

---

Feature status

Feature	Status	Task(s)
Multi-model routing (Grok / Gemini / Codex)	✅ implemented	models-001/002/005
Local model support — Ollama / vLLM	✅ implemented	models-003
Cost tracking and budget caps	✅ implemented	runtime-005
Session checkpoint / restore	✅ implemented	core-002/003
Audit log with hash chain	✅ implemented	enterprise-002
Team configuration and policy	✅ implemented	enterprise-001
Web search (debug / research)	✅ implemented	tools-003
Code index + whole-codebase mode	✅ implemented	memory-001/005
Git workflow automation	✅ implemented	autonomous-003
Verification loop	✅ implemented	autonomous-002
Self-healing debug loop	✅ implemented	autonomous-004
Multi-file orchestration	✅ implemented	autonomous-001
Impact analysis gate	✅ implemented	autonomous-005
Autonomy slider — 5 levels	✅ implemented	ide-004
VS Code extension	✅ implemented	ide-001/002/003
Voice interface (PTT + streaming)	✅ implemented	voice-001/002/003
Remote / LAN server + phone UI	✅ implemented	remote-001/002/003
Chrome extension + bridge	✅ implemented	extensions-002
MCP protocol server	✅ implemented	integrations-003
GitHub Actions CI integration	✅ implemented	integrations-001
Playwright browser automation	✅ implemented	tools-002
Sub-agent parallel execution	✅ implemented	tools-005
RBAC (role-based access control)	📋 planned	enterprise-003
Usage analytics dashboard	📋 planned	enterprise-004
Hermetic sandbox	✅ implemented	moonshot-002
Patch risk scoring	✅ implemented	moonshot-006
Impact graph smart test selection	✅ implemented	moonshot-007
CI eval pipeline	✅ implemented	evals-002
Continuous memory	✅ implemented	memory-003
Proxy/relay mode (geo-restricted regions)	✅ implemented	proxy-001
OIDC identity integration	📋 planned	—

---

Installation

Recommended (isolated environment):

pipx install gdmcode

pip:

pip install gdmcode

Development install:

git clone https://github.com/guidegdm/gdmcode
cd gdmcode
pip install -e ".[dev]"
gdm health

Note: The PyPI package is gdmcode. The command is gdm.

---

Configuration

gdm uses a 7-layer precedence system (highest wins):

CLI flags > env vars (GDM_*) > team policy > team preferences
         > project .gdm/config.toml > user ~/.config/gdm/config.toml > defaults

Quick setup:

gdm login grok        # or: gemini | codex | all

Secrets are stored in the OS keychain — never in TOML files.

Interactive setup is also available inside gdm with /login, /logout, and /proxy. First launch without credentials shows onboarding instead of starting model agents; type a normal message like hey to test the connected LLM.

Full configuration reference: docs/configuration.md

---

Enterprise

gdm includes audit and governance features suitable for team deployment:

• Audit log — every tool call is logged with a hash chain for tamper detection.
• Team policy — .gdm/team.toml enforces autonomy caps, network allowlists, and git permissions.
• RBAC — planned (enterprise-003).

See docs/security-hardening.md for deployment guidance.

---

Architecture

gdm runs a tool-calling agent loop: the model receives a task, selects tools from a registered registry, executes them (subject to injection checks, RBAC, and network policy), and writes the results back into context. Sessions are checkpointed to SQLite; memory is indexed for whole-codebase reasoning.

Full architecture diagram: docs/architecture.md

---

Contributing

See CONTRIBUTING.md. Issues and PRs welcome.

---

License

MIT — Copyright guidegdm