What changed
ConeCode has been released as a desktop coding workspace designed for AI agents, built using Electron, React, and TypeScript. This application integrates a suite of development tools, including chat interfaces, file management, a code editor, terminal sessions, Git worktree support, and a browser preview. A core distinguishing feature is its implementation of bounded, evidence-driven Recursive Self-Improvement (RSI), a mechanism for AI agents to iteratively improve software based on measurable outcomes.
The workspace offers four distinct agent modes: Standard for one-off tasks, Goal for long-term deliverables with external checks, Orchestrate for managing sub-agents, and RSI for measured self-improvement. ConeCode prioritizes project context, feeding agents information from files, Git changes, TODOs, and scripts to generate concrete task cards. Execution is bounded, with every file edit and command presented as an approvable action card, and a macOS Seatbelt sandbox limits command writes. Model management is robust, supporting multiple providers (OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter, custom endpoints, and a free ConeCode endpoint), with configurable pricing, context windows, and output limits.
The RSI mode is specifically designed for improving existing software against defined metrics like reliability, test health, or latency. It operates as a disciplined loop, where agents propose and implement small improvements within a declared scope. The process is governed by an external oracle that freezes goal vector metrics and baselines before an episode. A hard verification gate runs type checks, tests, and metric commands, with any failed criterion resulting in a hard failure. Each episode focuses on one module to maintain reviewability and facilitate rollbacks. Passing candidates are scored on alignment, simplicity, and discipline, with oversized diffs penalized. A real archive retains lineage, metrics, and rationale for accepted and rejected variants, allowing for retries of strong but underexplored branches. Slow registration of persistent learnings limits contamination from single successful runs, and stop conditions prevent unnecessary changes at plateaus or scope boundaries. Non-negotiable limits prevent RSI from weakening security, sandboxing, or tests.
Agent Modes
- Standard: Default coding loop for one-off tasks.
- Goal: Pursues declared goals until external completion checks pass.
- Orchestrate: Conductor agent decomposes work and spawns parallel sub-agents.
- RSI: Measured improvement loop for iterative, measured enhancement.
Workspace Features
- Slash Commands: /help, /clear, /compact, /plan, /review, /agents, /goal, /mode, /model, /memory, /worktree, /add-dir, /init, /diff, /commit, /image, /imagine.
- Custom Commands: Defined in /.conecode/commands/*.md.
- In-place Editing & Rollback: Edit past messages and restore revertable file changes.
- Handoff: Export transcript with takeover prompt for other AI tools.
- Project-Aware Suggestions: Reads local project facts for task proposals.
- Model Selector: Supports multiple providers, favorites, recent history, pricing, context/output limits, and a thinking-effort slider.
- Image Studio: Generates images with aspect ratio, quality, style controls, and art presets, reusing configured providers.
Why it matters for builders
ConeCode provides a unified desktop environment that consolidates essential development tools, reducing the friction of switching between applications for AI-assisted coding. The introduction of bounded RSI offers a novel and structured approach to AI-driven software improvement. Builders can leverage this to automate iterative enhancements, test hypotheses rigorously against defined metrics, and maintain a clear audit trail of changes, potentially accelerating development cycles and improving code quality in a controlled manner.
Practical impact
Developers can download and install ConeCode to experiment with its integrated coding environment. They can explore the different agent modes, particularly RSI, by defining specific measurable outcomes for their projects. The workspace's project-aware suggestions and bounded execution model allow for safe experimentation with AI-driven code modifications. Builders should review the .conecode/rsi/ directory to understand the RSI process and its outputs. Experimenting with custom commands and model configurations will also be key to tailoring the workspace to specific workflows.
Caveats and source limits
The source material describes ConeCode as a fresh release, with the latest version noted as v2.4.21. While it details the functionality and RSI mechanism extensively, specific performance benchmarks or independent validation of its RSI capabilities are not provided. Pricing details for the various model providers are mentioned as being visible within the application, but exact figures are not specified in the excerpt. The source also notes a macOS Seatbelt sandbox, implying platform-specific sandboxing capabilities, but comprehensive cross-platform security details are not elaborated upon.
Sources
Claim check: 8/8 supported claims - 8 evidence links - 100% avg confidence
- ConeCode is a desktop coding workspace for AI agents that integrates chat, files, editor, terminal, Git worktrees, and browser preview.supported - github.com
- ConeCode features bounded, evidence-driven Recursive Self-Improvement (RSI) for agents to improve software.supported - github.com
- ConeCode offers four agent modes: Standard, Goal, Orchestrate, and RSI.supported - github.com
- ConeCode supports multiple AI model providers including OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter, and custom endpoints.supported - github.com
- ConeCode includes an Image Studio for generating images with various controls and presets.supported - github.com
- ConeCode's RSI mode is designed for iterative, measured improvement of software against defined metrics.supported - github.com
- ConeCode's RSI operates with an external oracle for goal vectors and a hard verification gate for candidate changes.supported - github.com
- ConeCode's latest release is v2.4.21.supported - github.com
Caveats
- Single-source caution: verify critical details at the linked source.
Radar score 88/100 - how it was calculated
- Reliability 82: GitHub metadata supports source trust
- Freshness 100: Fresh GitHub release date
- Novelty 81: Fresh GitHub release
- Technical 85: Repository technical metadata
- Developer 96: Developer tooling signals
- Ecosystem 72: Fresh GitHub release
- Confidence 100: Claims have reliable evidence
Discussion
Loading comments...