For engineering teams

Every AI agent starts from zero. Yours won’t.

Imagine having your own AI model that knows your organization’s specific context and codebase. That’s Ryo. A one-of-a-kind custom model that plugs into your AI workflow to give frontier models the facts they’re missing. Built for fewer turns, fewer mistakes, and big savings.

Build your model Talk to us

How it works

Meet Ryo. Your own personal model that learns from your team’s work every night.

Learns only from approved work

Every night, Ryo learns from your team’s work: the pull requests you merged, once every check you run has passed. Work that failed its checks, or never merged, stays out.

#1491Move order list to cursor paginationmerged · checks passedLearned tonight

#1490Try a new retry libraryclosed · not mergedSkipped

#1489Fix claims field mappingmerged · checks failedSkipped

Built to keep what it knows

Most models lose old skills when they learn new ones. Ryo is designed not to.

/v2/claims replaces /v1learned Tuesday

paginate() for listslearned in March

toCents() for moneylearned last year

Works with

agent · acme/checkout

› Add retries to the claims sync job

reading src/claims/sync.ts

planning a change to 3 files

ryo /v1/claims was deprecated on Tuesday in #1423. Use /v2/claims — see src/claims/client.ts:88

✓ edited 3 files · tests pass

done in 11 turnslast similar task: 14

pull request #1488 · acme/checkout

Rryo reviewed #1488 · 1 note

src/orders/list.ts

- const page = rows.slice(offset, offset + 50)

+ const page = paginate(rows, cursor)

Pagination here doesn’t follow your team’s pattern. We use cursor pagination through paginate() — see src/lib/paginate.ts:12, used in 14 places.

What Ryo does for your agents

Knows your internal APIs.

Built to give the right endpoint and client before the first call, not after three failed ones.

Follows your organization’s patterns.

Retries, pagination, logging and money, done the way your team does them.

Remembers your fixes.

Built so a bug your team fixed once stays fixed.

Reviews every pull request.

Checked against how your team writes code, citing your own examples.

Why it matters

Your agents pay to rediscover what your team already knows.

66%

of developers name AI’s almost-right answers as a top frustration¹

1 in 2

AI fixes that pass their tests still wouldn’t be merged²

30×

token swing on the same agent task, with no gain in accuracy³

20%+

more cost from preloading instruction files, with no real gain in success⁴

The savings

Fewer turns, fewer repeated mistakes, smaller bills.

Agents fail when they don’t have the right context. Instead of preloading everything and racking up token cost, Ryo is built to deliver the facts that matter, when you need them. Built for better accuracy, fewer turns, and a smaller bill.

Savings by codebase size

250,0001 million5 million25 million

$20,000–$30,000

Estimated savings a year, assuming a 10–15% cut

Built for fewer repeated mistakes, too.

Your data

Your model is yours.

Your team’s work trains only your Ryo, and only your team can use it.

Let Airgetlam use your code to improve its own modelsOffNot available. Airgetlam doesn’t train its own models on any customer’s code. Your team’s work only ever trains your Ryo.

Your own model

Every company gets its own Ryo, trained on its own approved work and never on another company’s.

Your own encrypted storage

Your code and your Ryo are kept apart from every other customer’s, under a key of their own.

Never used to train ours

Airgetlam never trains its own models on your code.

You choose what it learns

Pick the repositories and paths, exclude the rest. Secrets files are always skipped.

Every change on record

Every settings change is logged with who made it and when, and so is each addition or removal of repositories — how many, never which.

Pricing

See what your agents are missing.

Start free. Then build your model.

Your Ryo

$249

per month

Learns up to 1 million lines

Your own model, learning every night from your team’s work
Review on every pull request
Search across your codebase or documents
Works with Claude, ChatGPT, Gemini and open-source models via MCP
Your own encrypted storage, with a key of its own
No per-seat fees, and as many repositories as your plan’s line limit allows
Roles and an audit log

Add BoostAdd-on+$300/moUpgrade to our largest model, built for your hardest work.

Build your model

Try Ryo Mini free for 7 days. Limit 250,000 lines.

Start free trial

Currently, Ryo is only available to US companies, teams, and users.

How we estimate savings. Estimate, not a measurement. Our model assumes agents take about 16% fewer turns, as reported for agent memory in published research,⁵ which cuts a coding task’s cost by about a fifth. It assumes coding is roughly 60% of token use, with search and support bots saving more and document work very little, and current list prices with prompt caching.⁶ It subtracts what built-in caching and context tools already save, for 15–25% fewer tokens and 10–15% lower spend. Each codebase size assumes one engineer for every 10,000 lines, at $150–250 per engineer per month, Anthropic’s published enterprise average for Claude Code.⁷ Real teams vary widely, especially teams that write code with AI, so read these figures as a guide, not a quote.

Ryo skips nights when little or nothing has been added or changed, like over a quiet weekend.

Sources

Stack Overflow, 2025 Developer Survey: AI (July 2025).
METR, Many SWE-bench-passing PRs would not be merged into main (March 2026).
Bai et al., University of Michigan, Stanford, MIT and others, How Do AI Agents Spend Your Money? (April 2026, preprint).
Gloaguen et al., ETH Zurich and LogicStar.ai, Evaluating AGENTS.md (2026, preprint).
Ouyang et al., Google Cloud AI Research, ReasoningBank (ICLR 2026).
Anthropic, Prompt caching (accessed September 2026).
Anthropic, Claude Code: Manage costs effectively (accessed September 2026).