Stop paying a frontier model to push to git.
Think locally. Reason globally.Socrates is a local-first layer for AI coding agents. Routine work runs on your own hardware. Claude, Codex or Gemini step in only when the problem actually needs them. Same results, a fraction of the frontier bill.
One email when there's something to try. Unsubscribe anytime.
Your AI bill isn't an intelligence problem. It's an efficiency problem.
Coding agents send almost everything to the most expensive model available. So the smartest model you pay for spends much of its day doing work a small model could do.
Searching the repo
Grepping, listing files and reading code that turns out to be irrelevant, all billed at frontier prices.
Routine operations
Commits, renames, formatting, boilerplate and one-line fixes that don't need deep reasoning.
Figuring out context
Burning tokens just to decide what to read before any real thinking starts.
You don't need to replace Claude or Codex. You need them to stop doing the easy parts.
Every task goes to the cheapest intelligence that can do it right.
Socrates splits each request into steps and routes each one across three levels of compute. It escalates only when a level can't handle the work, and hands the simple parts back down.
A small model on your machine
- Navigation and search
- Edits, refactors, renames
- git, tests, tooling
- Preparing context for the levels above
A larger model you already host
- Diagnosis and debugging
- Multi-file changes
- Work too heavy for a laptop
- Code that can't leave your network
Claude, Codex or Gemini
- Architecture and design
- Hard reasoning
- Problems the lower levels can't solve
- Delegates the follow-up work back down
Pre-processed context
The local level finds what matters, so frontier models read a focused brief instead of your whole repo.
Verified results
Changes are checked with your own tests and tools before they reach you. You get a working fix, not a pile of model replies.
Every dollar accounted for
See what ran where, why it escalated, and what it cost, per task and per week.
A receipt for every week of work.
Socrates tracks what it would have cost to send everything to a frontier model, and what you actually paid. Share it, show it to your manager, or just enjoy it.
// Example figures for illustration only.
If we can't prove it, we don't ship it.
Free for developers. Licensed for companies.
Free
- Full local + frontier routing
- Your personal savings receipt
- Local benchmarks and metrics
- Community on GitHub and Discord
Per-developer license
- SSO and user management
- Policies for what code may reach frontier models
- Cost dashboards by team and project
- Audit logs and compliance
- Managed level-2 server on your GPUs
- Support and SLA
The questions developers ask first.
Does my code leave my machine?
Only when a step is escalated to a frontier model, and only the context that step needs. Companies can set policies that keep specific code on local or self-hosted models. Socrates never uses your code to train anything.
Do I need a GPU?
No. Socrates detects your hardware and picks a local model that fits. Apple Silicon and machines with a recent GPU get the most out of the local level. If your machine is too slow for a step, Socrates escalates instead of making you wait.
Do I have to switch from Claude Code or Codex?
No. Socrates is designed to sit underneath the tools you already use, so your workflow stays the same and the bill gets smaller.
Will the cheaper models make mistakes?
Small models are only given work they can handle, and every change is verified with your tests and tools before you see it. Destructive operations always go through an extra check. When in doubt, Socrates escalates.
Is it really open source?
Yes. The core will be open source and free for individuals. Companies pay for the features only companies need, like SSO, policies and team dashboards.
When can I try it?
Socrates is in development. Join the waitlist and we'll let you know as soon as the alpha opens.
Get the same work done for a fraction of the bill.
Join the waitlist for early access and our benchmark results.