Runtime, memory, gateways, evals

BUILD INCLAUDECODEXOPENCODERUN IT ON OYA.

One platform under the agents your customers touch: runtime, memory, gateway connections, a skill library, secrets, evals and a single invoice. Bring the agent you already built, keep your tools, change what runs it.

  • fewer tokens per task
  • 100% state preserved byte for byte
  • 1 invoice, not 14
  • 41× cheaper in year one
  • 6.4× on recurring tooling alone
https://getoya.ai/api/mcp
Add as a custom connector, then sign in. No API key. Setup instructions

White paper · Documentation · Open source core · Pricing

Why Oya

The most efficient place to run an agent.

A token loop re-sends every tool result back through the model on the following step, spending tokens, latency and determinism to move state the model never needed to read. Oya compiles the plan once and wires each value by reference.

MeasureVercel AI SDKMastraoya
total tokens5,75424,3872,536
model round-trips662
latency17.6s15.3s4.7s
execution ordermodel-chosenmodel-chosenone fixed DAG
state fidelityunguaranteedunguaranteedguaranteed

Same task, same tools, same model: reconcile on claude-haiku-4-5, measured over 8 trials on the real API. The difference is who moves the data. Run the benchmark yourself

100%
Critical values preserved byte for byte
against 19 to 81% for published agent frameworks, on 120 projection-annotated tasks
Fewer tokens per task
4.7× as a geometric mean across both baselines above: 2.3× against the leanest, 9.6× against the heaviest
6 / 6
Frontier models, no state loss
the same result on every model tested, with the method set out in the white paper
The platform

What you need to build production agents.

Eight pieces of plumbing sit between a demo that works and an agent a customer pays for. None of them are your product. All of them are already here.

  1. Memory store

    Persistent state that lives outside the model context, so the goal doesn't drift on run two.

  2. Gateways and MCP

    One connection layer to the systems that matter: 32 gateways today, plus your own MCP servers.

  3. Versioned skills

    Typed, sandboxed functions with credential isolation. 96 in the catalog, versioned like code rather than pasted into a prompt.

  4. Guardrail harness

    Hard budget, step and scope limits. A runaway loop stops instead of quietly billing you.

  5. Structured planning

    One plan up front, then deterministic execution. The model decides intent, not order.

  6. Evals

    Replayable traces and scored cases, so "it seems better" becomes a number before you ship.

  7. State preservation

    The model works with a handle, not the contents. It cannot read the value, so it cannot corrupt it.

  8. Audit trail

    Every run recorded with inputs, tools, output and cost. Hand it to a customer, or an examiner.

Cost teardown

What one agent costs to build yourself.

An AI employee is not one purchase. It is a stack: a model to reason, a harness to run the loop, tracing to see why a run failed, search and a browser to research a prospect, a sandbox to run tool code, a messaging API, an OAuth layer, durable scheduling, memory, version control, a CI pipeline, and cloud infra under all of it. Then someone has to build the integrations, and keep building them. Here is the bill.

The agent being priced: Sends 20 personalized LinkedIn invites a day, researches every prospect before it writes, comments on peer posts, and routes inbound DMs to Slack.

Oya ships this one as a template. The teardown prices building the same thing from parts.

Build it yourself14 LINE ITEMS · PER MONTH
LLM inferenceAnthropic API$200
Agent harness and hostingLangGraph Platform$35
Tracing and observabilityLangfuse Pro$199
Evals and regression suitesLangSmith Plus, 1 seat$39
Web searchExa or Tavily$40
Browser automationBrowserbase Startup$99
Code sandboxE2B Pro$150
LinkedIn and messaging APIUnipile$53
OAuth and integrationsNango$250
Durable workflows and cronInngest Pro$75
Vector store and memoryPinecone Standard$50
Source control and CI/CDGitHub Team plus Actions$50
Cloud infraGKE and managed Postgres$150
AI coding subscriptionClaude Code Max 20x$200
Tooling subtotal, per month$1,590$19,080 per year
On Oya1 LINE ITEM · PER MONTH
All of it, included$250per live agent, per month
or $2,500 per agent, per year
  • Search, browser and sandbox built in
  • Run traces and cost accounting included
  • 32 gateways, OAuth handled
  • Routines, triggers and memory native
  • Specs versioned to your own GitHub repo
  • Live the same day
Who does the building

A senior engineer, or Oya Engineer.

Tooling is the cheap half. A Nango subscription gives you OAuth, not connectors: every integration is still code somebody writes, tests, versions and repairs when the API moves. This is the line most build-versus-buy models leave out, so here it is priced on both sides, along with the two figures that outlast the build: how long until the agent is doing the job, and what it costs to keep it doing the job.

Build it yourself
A senior engineer
one person writing the harness, the connectors, the versioning and the pipeline, then owning all of it
Time to first production run3 to 4 monthsharness, then connectors, then the pipeline, in sequence
Harness and infra wiring, 6 to 8 weeks$27,000
Building 8 integrations, about 3 days each$18,000
Spec versioning and CI pipeline$8,000
Maintenance, about 15% of an FTEevery year, forever$30,000
Year one, people$83,000
On Oya
Oya Engineer
an AI engineer inside the platform, with every built-in tool disabled so it can only act through Oya's own toolset
Time to first production runSame daydescribe the agent, review the spec it writes, deploy
Writes the agent from a specincluded
Authors and edits skillsincluded
96 skills and 32 gateways prebuiltincluded
Maintenance when an API moveswe ship the fix, not you$0
Year one, people$0
Year one, all in

$102,080 against $2,500.

Build it yourself · $19,080 tooling + $83,000 build and maintenance$102,080
Oya · one agent licence, billed annually$2,500
41×

41 times cheaper in year one, and about 6.4× on recurring tooling alone. The gap widens with the second agent: on Oya that is another $250 line, not another integration project.

Audit

What covers what.

Each line above, and the shipped Oya capability that stands in for it. Every entry maps to something already in the product, not to a roadmap.

Layer
Buy or build separately
On Oya
LLM inference
Anthropic API
Metered per run, or bring your own key across 8 providers
Harness
LangGraph Platform
The agent runtime: a spec compiled to Python and executed in a sandbox
Tracing
Langfuse Pro
Run traces with a per step trace graph, token and cost accounting included
Evals
LangSmith Plus
Built in eval runner against saved happy and edge cases
Web search
Exa or Tavily
web-search and fetch-url skills, metered per call
Browser
Browserbase
Oya Browser, a first class gateway and skill
Sandbox
E2B Pro
Isolated skill sandbox, billed per second, under a tenth of a cent per run
LinkedIn API
Unipile
linkedin, linkedin-api and linkedin-messaging skills
OAuth layer
Nango
32 gateways, 8 channels and 24 apps, tokens and refresh handled
The integrations
About 3 days of work each
96 skills in the catalog, or Oya Engineer authors a new one
Scheduling
Inngest Pro
Routines and triggers, cron and webhook, free to set up
Memory
Pinecone
memory and search-kb skills over a managed knowledge base
Versioning
GitHub plus a sync layer you write
GitHub sync: connect once, each agent gets a repo, every sync is a tracked commit
CI/CD
GitHub Actions minutes
No pipeline to run. Deploys are a platform action, and evals are the gate
Infra
GKE and managed Postgres
Managed. There is no cluster for you to run
The engineer
A senior engineer, year one
Oya Engineer, an AI that builds and edits agents through Oya’s own tools

Method and sources

  • Vendor prices. Published list rates as of August 2026, taken at the lowest tier that supports a production workload: Langfuse Pro $199, Nango paid entry $250, E2B Pro $150, Browserbase Startup $99, Inngest Pro $75, Pinecone Standard $50, LangSmith Plus $39 per seat, LangGraph Platform $35, Unipile from 49 euro per month, Claude Code Max 20x $200. Where a vendor publishes a range, the low end is used.
  • The CI/CD line. GitHub Team at $4 per user per month for 3 seats, plus about 5,000 Actions minutes beyond the 3,000 included at the $0.006 per minute Linux rate, plus container storage above the 2 GB allowance. Docker builds and a test suite on every push reach that volume quickly.
  • Usage-based lines. LLM inference, web search and cloud infra are estimates for one agent at the volume described, not list prices. Your numbers will move with volume.
  • Engineering cost. One senior engineer at a fully loaded US rate of about $95 per hour: 6 to 8 weeks to first production run, 8 integrations at roughly 3 days each, 2 weeks for spec versioning and a deploy pipeline, and maintenance at 15% of an FTE thereafter. Adjust the rate for your market and both totals move together.
  • Time to first production run. The three build lines run in sequence at one engineer, which is 3 to 4 months of calendar time and assumes nothing goes wrong. Same day on Oya is the deploy path the walkthroughs show end to end, not a best case: describe the agent, review the spec it writes, deploy. Neither figure includes the time it takes you to decide what the agent should do.
  • The maintenance line. It is called out on both sides because it is the largest single figure in the teardown and the only one that repeats every year: an integration is not finished when it ships, it is finished when the vendor stops changing the API. On Oya that work is ours, which is what the $0 means; it does not mean nothing breaks.
  • Oya Engineer. The platform’s own build agent. Every built-in tool is disabled, so it can only act through Oya’s own toolset: it writes the agent from a spec, edits skills, wires gateways, and runs evals against saved cases. It is not a replacement for engineering judgment, and someone still writes the spec and reviews what ships. The $0 is the absence of a build project, not the absence of a human.
  • Oya pricing. The per agent licence: $250 per live agent per month, or $2,500 per year. Drafts and offline agents are free. Pay as you go is also available, metered per run, which lands well below the licence for low volume agents.
  • Not a claim of feature parity. Each vendor above does more than the slice an agent needs, and Oya buys some of its own inputs. The comparison prices the slice a team would have to procure and assemble, not the whole of any product.
Apply

Bring us the agent you already built.

Six questions, about a minute. If it is a fit we will run your own workflow side by side with what you have today, and you keep the comparison either way. Prefer to talk first? Book time with a founder or email [email protected].

Application
01 / 06

What best describes your business?