OdinLabsTalk to Odin
Menu

Start with the work

What do you need
to move forward?

Pick the decision facing your team. Each path shows what to bring, what to check and when to pause.

Find a useful first workflow.

Choose one repeated task, its source and the person who checks the result.

Explore this starting point

Make agent work reviewable.

Bound a repository change and define the evidence before it runs.

Explore this starting point

Organize our AI tools.

Map connections, routing decisions and who reviews model usage.

Explore this starting point

Get answers we can check.

Start with an owned source, a supported question and one to refuse.

Explore this starting point

Already know the work? See how the products fit around it

Look inside
the approach.

Explore how the pieces can work together, then inspect the product examples. The demonstrations use fictional data.

How a project works with Odin, in three steps.

Start with a real piece of work.

A process that takes too long, a question people keep asking, a change you want to make. We can start with an Innovation Sprint, set it up with you, or train your team on the job.

Use the tools you already have.

Your model keys, your coding agents, your repositories and your documents, connected through Odin. Swap a model or add a source later without starting over.

Each project makes the next one easier.

Keep the evidence, the decisions and the context from each piece of work. Architecture rules, classifiers and tailored models can become parts your team can check and improve.

Stop explaining the same thing twice.

When an AI tool forgets everything between sessions, every new task starts with the same explanation and the same checking.

Odin keeps the reviewed context, the evidence and the connections, so the next task starts where the last one ended.

Your setup, your data, your decisions.
AI GatewayIllustrative product preview

Route with a reason.

Policy configured
Your toolsOne API connection
Odin AI GatewayRouting · policy · budgets
Quality profileAvailable route
Balanced profileSelected by policy
Cost profileAvailable route

The actual model and usage remain attributable.

Provider choices

Access follows your configuration.

  • Orq
  • EUrouter
  • Requesty
  • OpenRouter

Choose a routing priority

Match capability and cost to everyday work.

This fictional request has one eligible route: EUrouter · Mistral. Changing the priority keeps the same example model; it does not compare model performance.

One request, accounted for

Fictional request · Summarize a supplier brief

Attribution
Operations · supplier-demo key
Priority
Balanced
Configured route
EUrouter · Mistral
Served model
Mistral Medium 3.1 · example
Static example receipt
Look under the hood

Connect existing tools through an OpenAI-compatible API. The selected profile expresses a routing priority; configured capability, budget and residency policies constrain the route.

Provider choices are separate from routing profiles. This fictional request has one eligible route, so changing its priority keeps the same example model. Actual routes and models depend on configured access and policy. The receipt is not live usage or a claim about processing location.

Point your tools at one OpenAI-compatible API. Choose how requests are routed, set budgets and see who used which model.

Explore AI Gateway

What needs
to work better?

You don’t need to know our product names. Pick the job you need to do, then inspect the starting point and what it requires.

Connect our AI tools and understand model usage.

Start with AI Gateway

Point your tools at one OpenAI-compatible API. Choose how requests are routed, set budgets and see who used which model.

Before you start

You need an issued Gateway key and provider access for your environment. The guide explains the connection; it does not create an account.

Check your Gateway connection Discuss your setup
Give agents code work we can review.

Start with Harness

Give an agent a written task with clear limits. Guards block prohibited actions, and the result comes back with evidence for a person to review.

Before you start

Start with a repository and a bounded task. Tool compatibility and release authority are established during setup.

Plan your first bounded change Discuss your setup
See what our AI work needs next.

Start with Command Center

Projects, work orders, sessions and usage in one workspace. See what needs attention and check what has been done.

Before you start

The workspace and integrations are configured for your team. The website demonstration is illustrative.

Review your first project result Discuss your setup
Find answers in our own documents.

Start with Odin Next

Ask questions over the knowledge you choose. Answers show their sources, or say there is no grounded answer, and useful context is kept for the next piece of work.

Before you start

Your knowledge sources and access must be configured. The website example uses sample material, not your documents.

Check your first sourced answer Discuss your setup

Product access and ongoing model, hosting and support costs depend on your setup. Confirm these before committing. Use the setup checklist. Session prices are shown separately on the sprint page.

Four products.
Start with one.

Choose AI Gateway for model access, Harness for scoped code work, Command Center to oversee the work, or Odin Next for sourced answers.

AI Gateway

One connection to AI models.

Point your tools at one OpenAI-compatible API. Choose how requests are routed, set budgets and see who used which model.

Harness

Agents that work within limits.

Give an agent a written task with clear limits. Guards block prohibited actions, and the result comes back with evidence for a person to review.

Command Center

See and steer the work.

Projects, work orders, sessions and usage in one workspace. See what needs attention and check what has been done.

Odin Next

Answers from your own knowledge.

Ask questions over the knowledge you choose. Answers show their sources, or say there is no grounded answer, and useful context is kept for the next piece of work.

What teams
use it for.

Three common starting points, each with a clear result.

All your AI tools
through one connection.

Your tools AI Gateway Team visibility

Connect existing tools through an OpenAI-compatible API, with routing, budget checks against recorded spend and per-model usage in view.

Explore the workflow

Agents that change code
you can review.

Intent Scoped work Review

Give agents a scoped work order and guarded actions, and give your team the evidence for the next decision.

Explore the workflow

Answers that show
where they come from.

Knowledge Answer Context

Ask questions over your own documents and get answers with their sources, or a plain “no grounded answer” when the sources don’t say.

Explore the workflow

We set it up
with your team.

We work in your repositories and tools, on real work, and leave your team able to run it.

Work with Odin

Connect it to your systems.

Model access, agent workflows and knowledge, set up inside the tools you already use.

Teach the team.

Your people learn by doing the work with us, including how to review it.

Explore implementation & enablement

Some work takes
months, not days.

Legacy code that needs untangling, security work, experiments in robotics, vision models, machine learning and space. Our R&D lab takes it on with you, and you choose how public it is.

Work with our R&D lab
In the openPublic experiment, published results, open-source code.Behind closed doorsPrivate and confidential. Nothing is published without your say-so.

Controls built in,
not bolted on.

Keys stored as hashes. Budgets refuse requests once recorded spend reaches the limit. Usage recording is best-effort; failed writes are not counted. EU-only routing for chat completions when you need it. Agent work with an explicit scope and guarded actions. Answers that cite their sources, or say there is no grounded answer. We show you each control and agree the setup with you.

How we approach trust

Part of our R&D
is public.

We publish selected engineering experiments, results included, and open-source tools such as the Apache-2.0 Blueprint Conformance Engine.

Inside Odin Research
Declared intentExecutable rule
Blueprint Conformance Engine
Clean fixture · passViolating fixture · fail
Run it yourself: the fixtures are public.

New on the R&D floor.

The latest experiments, results and notes we have published.

  1. Field noteEXP 012: Can a reviewer talk itself out of its false alarms?No. Asked to name a concrete failing input for every finding or drop it, MiniMax M3.1 Flash kept 249 of 269 of its own findings. False alarms stayed at 20 and 22 of 30 clean changes, against a pre-registered bar of 2.
  2. Field noteA free model on our review benchOn a sealed corpus of 30 real defects and 30 clean changes, MiniMax M3.1 Flash's counts sat in Claude Opus's range at about 1.5 cents a review. No reviewer flags few enough clean changes to vote, and an agent-run reviewer can read its own answer key.
  3. Field noteCan a free model find real bugs in our private code?On bugs in our private fix history that can be found from the code alone, MiniMax M3.1 Flash's first guess landed within two lines of the fix 63% of the time, against 23% for our current default. Only about 1 in 10 of the small fixes we sampled could be found that way.
Everything on the R&D site

What would you
like to work on?

Tell us about it