Airuncode
Local-first coding agents on your keys or your AI plans
207 followers
Local-first coding agents on your keys or your AI plans
207 followers
Airuncode is a local-first runtime for coding agents. Bring your own API keys or your Claude and ChatGPT plans, run local models through Ollama, and pay providers directly with zero markup. Agents review PRs, deploy to Cloudflare and remember your codebase.
Free Options
Launch Team / Built With
Nice job on the launch! Can each agent in a parallel run use a different model at the same time, or do they all have to share one?
@colegawin Thanks Cole! Yes, each agent in a parallel run can use a completely different model simultaneously.
Since Airuncode acts as a local model-agnostic runtime, you aren't locked into a single provider or model per session. You can assign different models—such as pairing an Anthropic model (like Claude 5 Sonnet) with an OpenAI model (like gpt sol) or a local Ollama/LM Studio model—to run side by side, bring their own perspective to the debate loop, or execute distinct sub-tasks concurrently.
I'm Gustavo, the developer of AIRUNCODE.
I've been building a desktop AI coding environment around a question that keeps bothering me:
Why does changing the AI model often mean changing the entire development workflow?
You use Claude for architecture, Codex for implementation, maybe OpenRouter for cheaper models, and Ollama when local inference makes sense.
Each model has its strengths.
But your project doesn't suddenly become a different project because you switched providers.
That's why I'm building AIRUNCODE.
The idea is simple: the workspace stays, the models compete.
AIRUNCODE brings together:
- Multiple AI providers, including supported subscription integrations, APIs, OpenRouter and Ollama.
- Reusable project knowledge through a knowledge graph.
- Git, branches, pull requests and code review.
- Cloudflare deployment workflows.
- Remote access to your desktop development sessions.
I'm not trying to build another AI model.
I'm building the environment around the models, so developers can choose the intelligence they need without rebuilding their workflow every time.
And building that environment has meant working on everything around the product too.
Recently, I fixed an SEO issue affecting URL consistency across 72 pages of the AIRUNCODE website. Not exactly the glamorous side of AI development, but that's what turning software into a real product looks like.
I'm looking for developers willing to try AIRUNCODE and give technical feedback.
Especially if you're already paying for multiple coding AI tools and wondering why your workflow has to be scattered across all of them.
Try AIRUNCODE: https://airuncode.com/
I'd genuinely like to know:
If you could keep one thing consistent when switching between Claude, Codex and local models, what would it be?
Project context? Tools? ConversatI'm Gustavo, the developer of AIRUNCODE.
I've been building a desktop AI coding environment around a question that keeps bothering me:
Why does changing the AI model often mean changing the entire development workflow?
You use Claude for architecture, Codex for implementation, maybe OpenRouter for cheaper models, and Ollama when local inference makes sense.
Each model has its strengths.
But your project doesn't suddenly become a different project because you switched providers.
That's why I'm building AIRUNCODE.
The idea is simple: the workspace stays, the models compete.
AIRUNCODE brings together:
- Multiple AI providers, including supported subscription integrations, APIs, OpenRouter and Ollama.
- Reusable project knowledge through a knowledge graph.
- Git, branches, pull requests and code review.
- Cloudflare deployment workflows.
- Remote access to your desktop development sessions.
I'm not trying to build another AI model.
I'm building the environment around the models, so developers can choose the intelligence they need without rebuilding their workflow every time.
And building that environment has meant working on everything around the product too.
Recently, I fixed an SEO issue affecting URL consistency across 72 pages of the AIRUNCODE website. Not exactly the glamorous side of AI development, but that's what turning software into a real product looks like.
I'm looking for developers willing to try AIRUNCODE and give technical feedback.
Especially if you're already paying for multiple coding AI tools and wondering why your workflow has to be scattered across all of them.
Try AIRUNCODE: https://airuncode.com/
I'd genuinely like to know:
If you could keep one thing consistent when switching between Claude, Codex and local models, what would it be?
Project context? Tools? Conversation history? Git workflow? Something else?ion history? Git workflow? Something else?
Several agents in parallel on the same repo is the part I'd want to hear more about. The failure I keep hitting isn't bad code, it's a second correct implementation of something the repo already had, because each session starts from a blank slate and it reads fine in the diff. Does the debate happen after all of them have read what's already there, or are they arguing about approaches before anyone checked whether one exists? Self-healing on a failed test worries me for the same reason, green doesn't tell you the fix belonged in that layer.
@asadmalik901
Spot on. You hit the exact two failure modes that make most multi-agent setups frustrating in real-world codebases. Here is how we address both in Airuncode's architecture:
@dev_gustavosis The symbol map catches the easy case, a helper with a matching signature. The one that actually got me was a fetch wrapper plus two hand rolled retry loops in the same repo, and no AST index flags those as duplicates because they don't look alike, they just do the same job. Scope locking on self-heal is the part I'd pay for, that's the rule I keep failing to enforce on myself. Does the write lock hold when the real fix genuinely is one layer down, or can the agent escalate once it's traced the stack?
@asadmalik901 The symbol map is strictly structural (AST), so it won't catch the 'fetch wrapper vs retry loop' case you mentioned—that's semantic, not syntactic. That's a known gap we're working on.Regarding the write lock: It can escalate, but only if the agent explicitly traces the stack and proves the root cause is in a lower layer. It doesn't just silently fix the dependency; it has to request permission to expand the scope first. We definitely need to tighten that logic."