Latere builds applications and platform services for autonomous agents: software engineering and document work, on top of identity, a model gateway, sandboxes, Git hosting, storage, and orchestration.
Latere is Latin for "to be hidden." What is hidden is human intelligence. In increasingly autonomous systems, human judgment does not disappear. It recedes behind every layer of decision-making, invisible but indispensable. Latere exists to ensure that this hidden human intelligence remains present, remains effective, and is never engineered away.
We build toward agents that listen, understand context, and take initiative across tools, models, data, and environments. They carry work forward autonomously in the background, within agreed boundaries, and proactively stay aligned with people: sharing progress, explaining decisions, and seeking human judgment when it matters. Every system we ship follows one principle: human intelligence stays in the loop. Agents take initiative. Humans stay informed and in control.
One identity across five capabilities, each built on an open-source component.
Identity provides single sign-on through OpenID Connect, with federated login, organizations, teams, and token issuance. It powers one account across every Latere product, for people and agents alike.
Agents runs and orchestrates AI agents with scoped access, visibility into every action, and durable state to resume from. To build multi-agent systems in your own Go application, check the open source core: Topos.
Models puts one gateway in front of your model providers. Issue and revoke access without handing out provider keys, set spend limits, and track usage through one API. Check the open source core: Lux.
Environments manages environments for agents and workloads. Run commands, attach a terminal, move files in and out, and tear down an environment when you are done. Check the open source core: Cella.
Code hosts Git repositories over HTTPS and SSH. Give each project, sandbox, or agent run its own Git remote, with S3 compatible storage as the source of truth. Check the open source core: Origo, and origo-web for browsing its repositories.
Storage keeps files and workspaces durable for people, agents, and sandboxes, with version history and permissioned sharing over S3 compatible storage and Postgres. Check the open source core: Arca.
Developer documentation for the platform lives at platform.latere.ai.
Two capabilities for running models on compute you control, each built on an open-source core.
Model Serving deploys and operates open-weight models, from a single GPU host to a Kubernetes fleet. Freeze and verify model weights, start an inference engine, and expose OpenAI- and Anthropic-compatible endpoints with health checks. These endpoints can sit behind Lux for model routing and access control. Check the open source core: Fornax.
Model Execution runs open-weight LLMs in pure Go, with no Python or vendor runtime. Check the open source core: Forma.
ReplicHAI audits whether a research paper actually reproduces. Give it a paper: it finds what the authors released, implements and re-runs the work, records every decision the paper left unwritten, and returns a verdict you can take apart. What reproduced, what did not, and what could not be judged, component by component, published with the full transcript beside it. Finished audits are public in the registry.
Wallfacer is an AI engineering teammate that turns ideas into working software. Talk through a plan, watch it build, and stay in control at every step: chat for exploration, specs for design, tasks for parallel execution, and code for precise edits. Open source, runs on your own machine, bring any LLM provider.
Lectio turns any document into data. Submit almost any file, PDF, Word, spreadsheets, slides, images, or web pages, and get clean text or structured fields back, with every detail sourced to where it came from on the page. One request, predictable pricing, a full record.
Tools and libraries to build with the platform and beyond.
| Project | What it is |
|---|---|
| latere-cli | One binary for the platform: sandboxes, model access, files, agent sessions, and adversarial code review. |
| agent-skills | Reusable workflows for coding agents. The same spec and release process in Claude Code or Codex. |
| pay | Sell credit, hold a balance, and spend it, in Go: a processor-neutral payment port and a credit ledger. |
| service-template | Production template for a Go backend with a Bun + React frontend, tag-driven releases, and deploy evidence. |
| sandbox-images | Container images that run coding agents and computer-use sessions, under Cella or standalone with Docker or Podman. |
AI as an Infrastructure is our open book on AI as an infrastructure, from systems to agents: history, design decisions, and foundations, written for software engineers new to machine learning. Also in 中文.
- Human judgment is irreplaceable. AI can execute, even reason. But deciding what matters, what is worth doing, and what trade-offs are acceptable still requires a human in the loop.
- Transparency over magic. Every AI action should be visible, auditable, and reversible. AI that cannot be inspected cannot be trusted.
- Autonomy is a spectrum. Full AI autonomy and full manual control are both valid. Where to draw the line should always be a human decision.
- Build for the long term. AI capabilities will keep evolving, paradigms will keep shifting. But the principle that human judgment belongs in the loop will not.
Found a vulnerability in a Latere service or repository? Read the security policy and report it through GitHub private vulnerability reporting or to security@latere.ai. It covers scope, testing rules, response times, and safe harbor for good-faith research.
