ATLAS Open Source
One developer, one machine, open source. Every decision it makes is readable.
ATLAS Open source under AGPL-3.0
More from the models you already run
Inferstep is an inference-aware execution and verification runtime. It sits between your agent and your inference server, allocates compute by task difficulty, checks every attempt, and returns what held up.
Install ATLAS
curl -fsSL https://raw.githubusercontent.com/itigges22/ATLAS/main/scripts/atlas-bootstrap.sh | bash
It needs 16 GB of VRAM, around 20 GB of free disk, and Docker or Podman already installed. Single user, AGPL-3.0.
Inferstep running a task: planning, candidate generation, and each stage reported as it happens.
What happens after your agent writes the code
Seven stages between the request and the write. You still review and decide. More has been tried and rejected first.
- 01The proxy receives the agent's proposed change and records it as candidate zero, the baseline.atlas-proxy
- 02The scorer reads the serving model's hidden states and sets the candidate count and reasoning budget.Geometric Lens
- 03The pipeline generates the candidate implementations and passes each one to the sandbox.V3 pipeline
- 04The sandbox builds each candidate and runs the project's available checks and its own oracles.Sandbox
- 05The pipeline rejects failures, unresolved cross-file calls, and code that runs without doing the task.V3 pipeline
- 06If no candidate survives, the runtime repairs against the recorded failures and runs the suite again.V3 pipeline
- 07A candidate replaces the baseline only with authorization. Unauthorized writes carry no verification metadata.atlas-proxy
The usual answer is a bigger model
For most teams that stops at procurement. The source has to leave the building, and nobody can inspect what happened inside the call. Inference-time compute is the other route, on hardware you already own.
Evidence
The scorer sets the effort per task
It reads the serving model’s internal states and sets attempts and reasoning budget per task, so a one-line fix and a six-file refactor stop drawing the same effort. No second model, no external call.
Environment
Low user counts still commit the whole GPU
Loading a model commits its VRAM whether anyone uses it or not. At low user counts the memory is spoken for and the compute is not. Several attempts per request put that capacity to work.
Allocation
The loop finishes before you see it
Every attempt is built and run in a sandbox, against the project’s checks and the runtime’s own. Failures are rejected and repaired. What returns records how far it got.
Comparison
What you have now, and what is missing from it
A hosted API and a model on your own hardware differ in where the code ends up. Neither changes what happens after the answer arrives.
| What is being comparedOne property per row. | Hosted frontier API The vendor runs the model. Your request leaves the network. Why that matters | Self-served weights You run the model. Your request stays inside the network. Where it stops | ATLAS You run the model, and the runtime checks what it returns. How it works |
|---|---|---|---|
| Runs on infrastructure you operate | no | yes | yes |
| Source and prompts stay inside | no | yes | yes |
| No vendor-operated component | no | yes | yes |
| Runs the models you are permitted to run | no | yes | yes |
| Swap the model without changing the loop | no | yes | yes |
| Verification makes no external call | no | no | yes |
| One request produces several attempts | no | no | yes |
| Compute allocated by task difficulty | no | no | yes |
| Attempts executed against real checks | no | no | yes |
| Rejects code that does not do the task | no | no | yes |
| Rejects unresolved cross-file calls | no | no | yes |
| Repairs against failures, not a restart | no | no | yes |
| Records how far each attempt got | no | no | yes |
| Marks which output earned that record | no | no | yes |
| Keeps capacity busy at low concurrency | no | no | yes |
| Per-stage record of what was consumed | no | no | yes |
What people use it for
Four reasons teams install it. Most start with the first.
Capability
Get more out of a smaller model
Policy, residency or the hardware in the rack usually picks your model. Inference-time compute makes that one go further.
Read the solution ProvenanceKnow what was actually tested
Months later, someone asks what was verified. The record is still there.
Read the solution SovereigntyWork on code that cannot leave
Teams whose source cannot leave have gone without. Inferstep runs where the code is.
Read the solution AdoptionKeep the hardware busy at low user counts
Capacity sized for concurrency you do not have sits idle. Several attempts per request use it.
Read the solutionHow you can get it
ATLAS is single user and open source. Inferstep is the multi-tenant deployment: many users on one install, isolated, under one policy.
One developer
One team
Inferstep Team
One deployment for a team, with isolated tenants, shared serving, administration and model policy.
An organization
Inferstep Enterprise
Organization scale, with the identity, audit and continuity terms a regulated buyer needs.
Tell us what you are working with
You do not need this form to download it or read the docs.