Video summary
OmniRoute + OpenCode is INSANE (Why I Dropped Claude Code)
Main summary
Key takeaways
Key technological concepts & product features
Why Claude Code was dropped (main argument)
The creator did not switch because Claude Code became worse. Instead, two open-source tools can provide a similar “coding agent in your terminal” experience for much less/no cost.
The proposed replacement is:
- Open Code: the coding agent.
- OmniRoute (Omni route / “OmniRout”): the model gateway/router that feeds models to the agent.
Claude Code limitations (what still holds it back)
- Proprietary + locked model choice
- Described as closed-source and only runs Anthropic models, so you can’t swap the “brain.”
- Metered usage
- $20/month pro plan.
- A rolling 5-hour token window with a weekly cap layered on top.
- If the window is exhausted mid-task, users must wait or pay-as-you-go.
- The speaker frames this as the main pain point for “terminal all day” usage.
Open Code (coding agent) — features emphasized
- Open-source agent running in the terminal
- Reads the repository, creates a plan, edits files, runs commands, and checks results against command output.
- Model-agnostic
- Supports 75+ providers out of the box.
- Examples mentioned: Anthropic, OpenAI, Google, Grok, OpenRouter, and local via Ollama.
- Can change model per project.
- Integrated with compiler feedback
- Hooks into the language server so after edits it retrieves real compiler/errors and feeds them to the model.
- Goal: reduce “guessing” by making the model see exactly what the editor/compiler sees.
- Local session state & undo
- Sessions stored in a local SQLite file.
- Changes tracked as Git snapshots, enabling undo.
- MIT license
- Source is inspectable/forkable.
- Multiple operating modes
- Interactive terminal UI.
- A single-shot command suitable for CI.
- A headless server exposing the agent as an API.
- Adoption stats used as validation
- Mentions 160,000 GitHub stars and “~7.5 million developers/month” (as stated in subtitles).
OmniRoute (model gateway/router) — features emphasized
- Free router the user runs locally
- One OpenAI-compatible endpoint on localhost that the agent calls.
- Connects to 230+ providers; 50+ offer free tiers; 11 are free forever (no credit card).
- Fallback routing (resilience)
- If a provider is down, rate-limited, or slow, OmniRoute can reroute the same request to another provider “mid-request” so coding continues uninterrupted.
- The “never stop coding” tagline is highlighted.
- Token compression to reduce cost
- Uses two passes (RTK and Caveman) to strip 78–95% of tokens for “tool-heavy” traffic.
- Constraint claimed: structured JSON, URLs, and code remain byte-for-byte identical.
- It compresses repetitive prose/tool chatter while preserving the actual source being edited.
- Privacy
- Runs entirely on the user’s own hardware; provider keys stay local; it “does not phone out.”
- Not just a proxy
- Mentions ~94 built-in tools via the Model Context Protocol.
- Includes a semantic cache to reuse previously computed answers without hitting the model again.
- Routing strategies
- 18 routing strategies, including:
- cheapest-first / fastest-first
- weighted splitting
- hard cost caps, etc.
- 18 routing strategies, including:
How the two are combined (tutorial/guidance described)
- OmniRoute includes a button/config generator that:
- Produces an Open Code configuration file listing all models OmniRoute can reach.
- Setup flow described:
- Generate the config from OmniRoute.
- Place the config file into the Open Code setup.
- Point Open Code to the local OmniRoute endpoint.
- Result: Open Code operates like Claude Code’s “read/edit/run/fix loop,” but with user-chosen models and fallback/free-tier routing.
Review-style conclusions / “honest places it still falls short”
- Free tiers have tradeoffs
- Slower models, tighter rate limits, and quality variance depending on which provider you land on.
- Fallback helps avoid interruptions but does not remove quality swings completely.
- User burden
- Claude Code is “install and works.”
- This setup requires more manual integration: two tools + config + managing provider keys.
- High-end tasks
- For “truly hard problems,” a top closed model like Opus still outperforms the free pack.
- Recommendation: use the best paid model only for the tasks that truly need it.
- Main claimed benefit
- “Agent and model finally came apart”: the coding loop stays open, the model becomes selectable, and cost becomes more optional.
- The speaker emphasizes they no longer hit Claude’s 5-hour meter because OmniRoute’s fallback keeps the session going.
Main speakers / sources
- Speaker: The YouTube video narrator/author (creator making the comparison and recommending OmniRoute + Open Code).
- Primary software sources referenced:
- Open Code (coding agent; associated with the SST team / now called Anomaly)
- OmniRoute / Omni route (local routing gateway)
- Claude Code (proprietary baseline being compared against)
- Mentions of Anthropic Opus as a high-end closed model option