Ogcode

What does this agent do, and when should you use it?

Ogcode is an open-source (AGPL-3.0, with a commercial license option), local-first AI coding agent and agent operating environment written in Go, operated through a browser interface served at http://localhost:9595. Before touching code, it maps the project and reads files by exact line ranges, keeping context tight and token cost low. It implements features across files, runs builds and tests, and investigates failures instead of retrying blindly, and can break objectives into tasks, run independent ones in parallel, and open pull requests for review. It pairs a memory system that recalls past decisions with reusable SKILL.md workflows, external MCP servers, built-in web search, and rich output including Mermaid diagrams, LaTeX, Plotly charts, and sandboxed HTML. Models are configured via environment variables or ogcode.json, supporting Anthropic, OpenAI, OpenRouter, and key-free local models through Ollama. Permissions are chosen per session — Ask, Auto, or Yolo — and deployment options include Docker plus documented remote-access boundaries such as SSH tunnels, authenticated reverse proxies, and VPNs.

Running ogcode serves the web UI at http://localhost:9595. The workflow: it first maps the project structure, then selectively reads files by exact line ranges to keep context focused; it then edits code across files, executes build and test commands, and investigates failures when they occur. It can turn an objective into tasks, run independent tasks in parallel, and open pull requests for review. Its memory recalls past decisions instead of replaying the full transcript, keeping long sessions affordable. It performs built-in web search and page reading, executes reusable SKILL.md skill workflows, connects to external MCP servers, and renders Mermaid diagrams, LaTeX, PDFs, Plotly charts, and sandboxed HTML with downloadable artifacts. Providers are detected from environment variables (ANTHROPIC_API_KEY, OPENAI_API_KEY, OPENROUTER_API_KEY, OLLAMA_BASE_URL) or from ogcode.json / ~/.config/ogcode/config.json, with environment variables taking precedence. Every tool execution is gated by a per-session permission mode: Ask (approve each step), Auto (risk-gated), or Yolo (no prompts).

1. A developer onboarding onto an unfamiliar codebase: have Ogcode map the project and outline files with exact line ranges instead of manually reading everything.
2. An engineer delivering a whole feature: split the objective into tasks, run independent ones in parallel, and open pull requests for review.
3. A privacy-conscious team: run with Ollama so code and session data stay local with no API key.
4. A developer debugging flaky builds: let the agent run tests and investigate failures rather than retrying blindly.
5. A user on a remote server: reach a deployed instance through an SSH tunnel, authenticated reverse proxy, or VPN from the browser.
6. A team standardizing workflows: capture repeatable processes as SKILL.md files with per-skill permissions (e.g. deploy-prod set to ask).

How do you install or deploy this agent?

One-line install script:

curl -fsSL https://ogcode.in/install.sh | sh

Or via package managers:

brew tap prasenjeet-symon/tap && brew install ogcode

winget install prasenjeet-symon.ogcode

On Windows without winget:

irm https://ogcode.in/install.ps1 | iex

On any OS with Go:

go install github.com/prasenjeet-symon/ogcode@latest

Building from source requires CGO_ENABLED=1 (the Swift tree-sitter binding is cgo). Alternatively run the container:

docker run -p 9595:9595 \
  -v ~/.ogcode:/root/.ogcode \
  -v "$(pwd):/workspace" -w /workspace \
  ghcr.io/prasenjeet-symon/ogcode:latest

The image is also on Docker Hub as prasenjeetsimon/ogcode:latest.

How do you use this agent?

Set at least one provider environment variable:

export ANTHROPIC_API_KEY=sk-ant-...

Or use a local model with no API key:

ollama serve
ogcode

Start it and open the web interface:

ogcode            # serves the UI on http://localhost:9595

Project settings live in ogcode.json at the project root, global settings in ~/.config/ogcode/config.json; environment variables override both:

{
  "providers": {
    "ollama": { "baseUrl": "http://localhost:11434" },
    "anthropic": { "apiKey": "sk-ant-..." }
  },
  "skills": {
    "paths": ["./team-skills"],
    "permissions": { "deploy-prod": "ask" }
  }
}

Pick a permission mode per session: Ask (approve each step), Auto (risk-gated), or Yolo (no prompts). For remote deployment, never expose the server unauthenticated; a recommended boundary is an SSH tunnel:

ssh -L 9595:localhost:9595 user@your-server

OG Lab's OGX subscription plan can also be connected from the settings screen; its models run through OG Lab's gateway with no environment variable needed.

What are this agent's strengths and limitations?

How does this agent compare with similar options?

Key facts side by side with the most closely related agents.

How does FollowAgents rate this agent?

Recent releases

Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.

1. v0.48.0 2026-10-17
   Release notes excerpt

   Release Notes — v0.48.0 Minor: The interface comes in light and dark Ogcode now has a light interface alongside the dark one it has always worn, and a Theme control in Settings → General to choose between them. The three options are Light, Dark, and System; System follows your operating system and switches live as your machine does, so a laptop that dims itself in the evening takes the app with it. Whatever you pick applies to the whole machine — every browser tab and every window — because the choice is kept by the server, not by one tab. The …

   View on GitHub ↗
2. v0.47.4 2026-10-17
   Release notes excerpt

   Release Notes — v0.47.4 Patch: Compaction keeps the prompt you started the turn with When a turn folded its context with compact_context, the message the turn opened with — what you asked for, and any image you attached — could fall out of the model's view. The summary took its place. Everything the agent did after that point was working from a digest of your request instead of the request itself. The fold now keeps your original prompt, images included, at the head of what is sent, with the summary folded in beneath it. If the turn is compacted again …

   View on GitHub ↗
3. v0.47.3 2026-10-16
   Release notes excerpt

   Release Notes — v0.47.3 Patch: Browser-backed search finishes what it starts Deep research fetches its chosen pages under one shared time budget, and the browser fallback was landing on the wrong side of it: the budget expired at the same instant an ordinary HTTP fetch timed out, so a page that needed a browser to render never got fetched. The window is now wide enough for the fallback to render and extract the page. On macOS that fallback is headless Chromium — Chrome, Chromium, Brave, or Edge, whichever is installed — with Safari kept only for a machine that has none of …

   View on GitHub ↗

FAQ