deepcode is an open agentic coding system that provides deep context, supervised execution, evidence-driven completion, and durable local work sessions for AI-driven development. It orchestrates multi-step tasks including real execution and iterative verification to translate natural-language goals and research papers into production-ready code.
Project overview
deepcode demands that task completion be supported by actual artifacts such as reviewable diffs, command output, and test results, feeding verification failures back into the task instead of stopping at a convincing-looking response.
Deployment
Refer to project documentation
License
MIT
Best for
Developers who need an agentic coding system that provides supervised execution, deep context, and evidence-driven completion for AI-driven development.
Researchers or developers who want to convert research papers into executable code through a multi-agent architecture with iterative verification.
Developers who need durable, resumable local work sessions indexed by workspace and discoverable across directories.
Key capabilities
Turns model responses into supervised work with native read, search, edit, shell, and test capabilities, keeping actions inside an execution contract with explicit permissions.
Allows the agent to keep working across turns against a mutable natural-language goal, deciding whether to continue, complete, or report a blocker based on conversation and tool results.
Builds context from the repository combining files, search results, project instructions, selected skills, and session history, compacting long histories against the active model's context window.
Demands that task completion be supported by actual artifacts like reviewable diffs, command output, test results, or recorded artifacts, feeding verification failures back into the task.
Stores sessions locally indexed by their original workspace, allowing them to remain discoverable across directories and resumable after restarts.
Allows users to configure named LLM connections and manage reusable workflow instructions (Skills) kept with a project or user library.
Delegates bounded tasks to agents in isolated Git worktrees and integrates their changes with conflict detection, supporting recurring automated work.
Multi-agent architecture that converts research papers into executable code through document parsing, intent understanding, code planning, indexing, generation, and iterative verification.
Limitations and risks
deepcode requires a Python 3.12+ environment.
An external or local LLM provider is required; deepcode depends on a configured LLM connection to operate.
Paid LLM services are optional; users may choose free or local providers, but a paid service may be needed depending on the selected provider.
External network calls are optional; selecting a remote LLM provider introduces external data transfer, while a local provider such as Ollama keeps inference local.
GPU requirements are not documented.
Telemetry practices are not documented.
Getting started
Installation is straightforward using a two-step process: install uv tool, then install deepcode and initialize.