Codel

What It Does

Codel is a self-hosted autonomous AI coding agent that runs entirely inside Docker. Users submit tasks through a browser-based web UI; the agent then autonomously plans and executes steps using three built-in tools: a terminal for running shell commands, a browser (powered by go-rod) for web lookups, and a file editor for viewing and modifying code. All execution history and command outputs are stored in a PostgreSQL database for persistent review. The backend is written in Go; the frontend in TypeScript.

The project launched in March 2024 and briefly attracted attention as one of the first Docker-native autonomous agent implementations with a polished UI. Development stalled at v0.2.2 (April 2024) and has not kept pace with the fast-moving autonomous coding agent landscape.

Key Features

• Autonomous task execution loop: terminal + browser + editor without human checkpointing
• Docker-based sandbox isolates agent actions from the host (via nested container creation)
• go-rod browser automation for real-time web information retrieval during task execution
• Built-in file editor displays modified files in the web UI as the agent works
• PostgreSQL-backed persistence stores full command history and outputs across sessions
• OpenAI support (default: gpt-4-0125-preview) with configurable model and endpoint
• Ollama integration for local/self-hosted model usage via OLLAMA_MODEL and OLLAMA_SERVER_URL
• Single docker run deployment with environment variable configuration
• AGPL-3.0 license ensuring all modifications must be open-sourced

Use Cases

• Use case 1: Local experimentation with the autonomous agent-in-Docker pattern on personal development tasks
• Use case 2: Reference implementation for studying the architecture of Docker-native coding agents (terminal + browser + editor triad)
• Use case 3: Privacy-sensitive or air-gapped environments where self-hosted LLM via Ollama is required and task complexity is modest

Adoption Level Analysis

Small teams (<20 engineers): Possible for individual experimentation. Setup is a single Docker command. However, stalled development, no benchmark data, and the Docker socket security issue make it a poor choice even for small teams with any production intent. Better alternatives (OpenHands, OpenCode) are more actively maintained.

Medium orgs (20-200 engineers): Does not fit. No multi-user support, no API, no integrations with issue trackers or CI/CD. The project is effectively unmaintained.

Enterprise (200+ engineers): Does not fit. AGPL-3.0 licensing alone is a blocker for many enterprise legal teams, and the project lacks any enterprise-oriented features (RBAC, audit logging, SSO, team management).

Alternatives

Evidence & Sources

• Codel GitHub repository — primary source, 2.4k stars, 202 forks
• go-rod browser automation library — underlying browser tooling
• Docker socket security implications — explains risks of mounting /var/run/docker.sock
• SWE-bench Verified Leaderboard (Epoch AI) — benchmarks Codel is absent from

Notes & Caveats

• Stalled development: Last release v0.2.2 was April 2024. The project has not been updated to support newer model APIs (GPT-4o, Claude, Gemini) or modern agent patterns. This is a significant gap given how fast the space evolved in 2024-2026.
• Docker socket security: The required --volume /var/run/docker.sock:/var/run/docker.sock mount grants the agent container effective root access to the host. This is a well-known Docker security anti-pattern. Purpose-built agent sandboxes (E2B, Microsandbox) avoid this via Firecracker or gVisor-based isolation.
• AGPL-3.0 licensing: Any software that incorporates Codel’s code or runs it as a networked service must release all modifications under AGPL-3.0. This is a practical blocker for commercial use cases.
• No benchmarks published: Unlike all major 2025-2026 autonomous coding agents, Codel has no published SWE-bench, HumanEval, or equivalent evaluation. Performance on complex tasks is unverifiable.
• Local model quality: Ollama support was designed for llama2-era models. Performance on autonomous coding tasks with llama2-class models is known to be poor industry-wide. The path is architecturally available but not practically useful for complex work.
• Historical value: Codel is a useful reference for understanding the early Docker-native autonomous agent architecture that OpenHands and others later built upon and refined.