InsForge: Open-Source Backend Platform for AI Coding Agents

InsForge (YC P26) is an open-source backend platform designed specifically for AI coding agents. Instead of agents juggling dashboards and manual config, InsForge gives them a CLI and skills to handle all backend/infra tasks end-to-end. The project is available on GitHub under Apache 2.0.
How It Works
Coding agents interact with InsForge through two interfaces:
- MCP Server (self-hosted and cloud): exposes InsForge operations as tools for any MCP-compatible agent.
- CLI + Skills (cloud only): a command line interface paired with Skills that agents invoke directly from the terminal.
Both allow agents to read backend context (docs, schemas, logs) and configure primitives (deploy edge functions, run migrations, create buckets, set up auth providers).
Core Products
- Authentication – user management, auth, sessions
- Database – Postgres relational database
- Storage – S3 compatible file storage
- Model Gateway – OpenAI compatible API across multiple LLM providers
- Edge Functions – serverless code on the edge
- Compute (private preview) – long-running container services
- Site Deployment – build and deploy sites
Agent-Specific Features
- Backend Branching: branch entire backend (DB, auth, storage, functions, schedules). Agents work on a branch, you review diffs, merge or discard. Inspired by Neon.
- Server Telemetry: agents read logs, CPU, memory, disk to find spikes and root causes.
- Debug Agent: every project gets a dedicated debug agent. Your coding agent asks questions like “why deployment fail?”, the debug agent runs diagnoses and proposes fixes.
- Backend Advisor: scans your backend daily for security and performance issues, proposes remediations, sends to your coding agent.
Quickstart (Self-Hosted)
git clone https://github.com/InsForge/InsForge.git
cd insforge
cp .env.example .env
docker compose -f docker-compose.prod.yml upThen open http://localhost:7130 to connect the MCP server. After setup, ask your agent: "I'm using InsForge as my backend platform."
Cloud version is available at insforge.dev.
The team notes they started InsForge because manual MCP setups had issues: tools pre-loaded into context, large payloads (>10k tokens), and missing capabilities like telemetry and configs. By putting everything into a CLI and teaching agents via skills, they avoided those problems.
📖 Read the full source: HN AI Agents
👀 See Also

TradingAgents-GUI: Local Web Interface for Multi-Agent Stock Analysis, Now with Ollama Support
A forked GUI for the TradingAgents multi-agent LLM stock analysis framework. Runs locally with Ollama, OpenAI, Anthropic, and more. Features live pipeline visualization, report reader, and token-saving concise mode.

Session Inspector for Claude Code provides real-time visibility into AI agent operations
Vibeyard, an open-source terminal IDE that wraps Claude Code, has added a Session Inspector feature that provides real-time visibility into Claude Code sessions with timeline tracking, cost breakdowns, tool analytics, and context window monitoring.

Multi-Agent Haiku System Matches Claude Opus on Complex Number Theory Problem at 15x Lower Cost
A Reddit experiment shows a two-Haiku agent system (generator + auditor) achieving identical 4/4 scores to Claude Opus 4.5 on a difficult Fermat's Little Theorem proof, while costing approximately $0.004 per query versus $0.06 for Opus.

Claude Agent Teams UI: Desktop App for Visualizing Claude Code Agent Workflows
A developer built a free, open-source desktop app that adds a visual layer to Claude Code's experimental Agent Teams feature. The app provides a real-time kanban board where tasks move automatically as agents work, plus cross-team communication, built-in review workflows, and per-task code review.