# Azumo — Nearshore AI development agency building agents, RAG and custom AI software, priced per Delivery Unit.

> Source: The Agents Index — https://theagentsindex.com/azumo (structured, researched, re-verified)
> Facts last verified: 2026-10-20

Azumo is a San Francisco-based, SOC 2 certified nearshore software and AI development company that scopes, builds, ships and operates custom AI systems. It also runs Valkyrie, its own hosted open-weight model service.

| Fact | Value |
| --- | --- |
| Website | https://azumo.com |
| Price | from $499/delivery unit pay-as-you-go |
| Free tier | No |

No free or trial tier on the published rate card; the only unpaid step is a scoping call or a paid discovery engagement, and Valkyrie publishes a login but no free plan.

## What it does

**Writes code** · Yes — Writing new software from a specification is the core of the engagement, and Azumo says coding assistants do the authoring while its engineers direct and judge the output.

> Code AI writes. Engineers guide and judge.
> — https://azumo.com · "AI Native Development" delivery-steps strip, Code step

**Debugs and fixes** · Partly — Azumo takes on maintenance and support of existing applications and runs a daily automated audit that reports severity, file and line, but its pages describe finding and grading defects rather than a guaranteed fix cycle.

> Ongoing maintenance, monitoring and support, with a daily automated code audit covering security, cost and architecture.
> — https://azumo.com/legacy-modernization · "What we do" section, Operate step

**Writes tests** · Partly — Verification is a named stage of their delivery process and they run application security testing of code, but Azumo's pages describe testing the work rather than writing a test suite as a deliverable.

> We perform static and dynamic software application security testing of code, including open source libraries, as part of our software development process when agreed to by our customers.
> — https://azumo.com/resources/security · "Secure Testing" section, third bullet

**Reviews code** · Yes — Every change their assistants produce is put through a code-auditing tool Azumo built itself, re-run daily across the whole codebase with graded findings including severity, file and line number.

> Every change our code assistants write runs through our code-auditing tool we built and use on our own products. It checks for security, maintainability, and the durability that keeps software healthy long after launch.
> — https://azumo.com · "Automated Code Review and Documentation" section

**Whole-codebase work** · Yes — Repository-wide migration is a service line: they modernise .NET and Azure estates in place across five named paths, in approved increments on the client's own repositories.

> Migration to ASP.NET Core, re-architecture, cloud migration and data platform work. Delivered in sprints, on your repositories.
> — https://azumo.com/legacy-modernization · "What we do" section, Modernize step

**Ships it** · Yes — Deployment is inside the engagement rather than handed back as a diff: they integrate into systems of record, get it live and then keep running it in production.

> Ship We integrate into your systems of record and get it live in production.
> — https://azumo.com/artificial-intelligence · "AI Development Services" four-step strip, Ship step

**Done-for-you service** · Yes — This is a client-services business: Azumo's own engineers scope, build, ship and then operate the AI system for you, as pods, staff augmentation or a full project build rather than software you license.

> One team can own the whole path. The same team of AI augmented engineers who scope your project are the ones who can build, ship, and operate it.
> — https://azumo.com/artificial-intelligence · "AI Development Services" section, "We build AI solutions and launch them into production"

## How much it does unattended

**Runs autonomously** · Partly — Azumo builds agents it describes as autonomous and 24/7, but its own published practice keeps writes to systems of record behind a human approval step, so unattended completion stops at the approval boundary.

> Write access to systems of record stays behind a human approval step, always — an agent that files its own conclusions is how bad data enters a business.
> — https://azumo.com/artificial-intelligence · "Research and Competitive Intelligence" use case, body paragraph

**Multi-agent** · Partly — They describe agents carrying one process across several systems and name orchestration frameworks they build on, but no page explains how multiple agents coordinate on a single task.

> Agents that carry a process across the tools it actually spans, taking actions and stopping where a human decision is required.
> — https://azumo.com/artificial-intelligence · "Workflow Automation Across Systems" use case, subheading

**Agent permissions** · Yes — Permission boundaries are an explicit design commitment: which steps an agent may execute and which need approval is decided before build, with action logging, rollback paths and rate limits on writes.

> We map the process first, decide explicitly which steps an agent may execute and which need approval, then build inside those limits.
> — https://azumo.com/artificial-intelligence · "Workflow Automation Across Systems" use case, body paragraph

## Which models it runs on

**Claude** · Yes — Anthropic is one of the frontier vendors they build on, their CTO is described as a Certified Claude Architect, and Claude Code is named as a daily tool and a delivered rebuild.

> We stay independent at the model layer choosing the best fit frontier model among OpenAI, Anthropic, and Google Gemini and the open weight alternatives like Qwen, Mistral, and Deepseek.
> — https://azumo.com/artificial-intelligence · "A vendor neutral approach" paragraph under Models & Platforms

**GPT** · Yes — OpenAI is named first among the frontier models they select from, and GPT-4o Realtime appears in the stack behind their own voice-agent work.

> We stay independent at the model layer choosing the best fit frontier model among OpenAI, Anthropic, and Google Gemini and the open weight alternatives like Qwen, Mistral, and Deepseek.
> — https://azumo.com/artificial-intelligence · "A vendor neutral approach" paragraph under Models & Platforms

**Open models** · Yes — Open weights are a declared speciality: they build on Qwen, Mistral and DeepSeek for clients and operate Valkyrie, their own service for fine-tuning and serving open-weight models.

> Valkyrie enables you to run open-weight AI agents fine-tuned on your data. You control the data and get predictable, consistent performance.
> — https://azumo.com/valkyrie · Hero section, paragraph under "The Best Models are Domain Specific and Purpose Driven"

**Model choice** · Yes — Vendor neutrality at the model layer is their stated position, choosing per use case across frontier and open-weight models, and on Valkyrie the customer picks the open-weight model and pins its version.

> We stay independent at the model layer. Our AI development services delivering using Claude, Codex, Qwen and others when they are the right tool.
> — https://azumo.com · Model logo strip paragraph, "we use Coding assistants" section

## Where you use it

**In your editor** · Partly — There is no Azumo IDE extension; the editor surface belongs to Valkyrie, which registers as a custom provider inside coding tools such as Cursor so engineers keep their own editor.

> Claude Code, Cursor, Opencode, Pi, and Qwen Code all connect to a Valkyrie deployment. Claude Code takes the Anthropic protocol directly. The rest register Valkyrie as a custom OpenAI-compatible provider.
> — https://azumo.com/valkyrie · "Access Valkyrie Your Way" section, Coding Agents card

**On the command line** · Partly — A terminal path exists only through Valkyrie, which CLI coding agents reach by pointing two environment variables at a deployment endpoint; Azumo's services themselves have no CLI.

> OpenAI and Anthropic compatible endpoints. Most tools connect with one environment variable and no code changes.
> — https://azumo.com/valkyrie · "How Valkyrie Works" section, Connect step

**In a browser** · Partly — The services engagement has no customer-facing web app, but Valkyrie ships a hosted dashboard for deployments, training jobs, usage and keys, with a public login.

> Deploy a model from Hugging Face or one you have fine-tuned, watch training jobs, read per-request usage, and manage keys and balance.
> — https://azumo.com/valkyrie · "Access Valkyrie Your Way" section, Valkyrie Dashboard card

## Whose machine it runs on

**Self-hosted** · Partly — Valkyrie itself runs on hardware Azumo operates, but fine-tuned weights are the customer's to export and run elsewhere, and client builds are deployed into the customer's own AWS, Azure or GCP account.

> A model we fine-tune on your data is yours. Export it and run it elsewhere whenever you want. It also means we are not a single point of failure.
> — https://azumo.com/valkyrie · "How your data is handled" section, "You own the weights"

## What it costs to run

**How it meters** · Yes — Client work is priced per Delivery Unit, a defined testable feature graded for complexity, with time-and-materials billed weekly as the alternative; Valkyrie instead sells concurrency per developer.

> A Delivery Unit is a defined, testable piece of working software. Essentially a working feature or sub-feature with its own acceptance criteria.
> — https://azumo.com/get-pricing · "WHAT IS A DELIVERY UNIT?" block below the rate cards

**Free tier** · No — The published rate card runs Starter, Growth, Platform and Custom with no free option, and the only free step before paying is a discovery engagement or estimate call.

> Software and AI development at $499–$899 per Delivery Unit.
> — https://azumo.com/get-pricing · Hero, "Custom Development Services Pricing Made for Outcomes"

**API access** · Yes — Valkyrie exposes OpenAI- and Anthropic-compatible REST endpoints, so an existing SDK or app connects to a deployment without code changes, which is the programmatic route into Azumo's own platform.

> Built on the OpenAI standard, so your SDKs and apps connect as they are. Point them at your deployment and go.
> — https://azumo.com/valkyrie · "Access Valkyrie Your Way" section, OpenAI Client card

**MCP server** · Partly — Valkyrie lists an MCP server as one of four access routes, but the card repeats the dashboard's description verbatim and no docs page describes its tools, so the scope of what an agent can query is unestablished.

> Deploy a model from Hugging Face or one you have fine-tuned, watch training jobs, read per-request usage, and manage keys and balance.
> — https://azumo.com/valkyrie · "Access Valkyrie Your Way" section, MCP Server card

**Bring your own key** · Partly — Client systems are deployed into the customer's own cloud account and Valkyrie weights are exportable, but no page offers bringing your own model API keys or your own GPUs to Valkyrie.

> We deploy on AWS, Azure, and Google Cloud.
> — https://azumo.com/artificial-intelligence · "A vendor neutral approach" paragraph under Models & Platforms

## Buying it for a team

**A company can buy it** · Yes — The buyer is an organisation throughout: enterprise programmes and multi-phase platforms are a named price tier and the stated customer base runs from startups to the Fortune 100.

> Platform $499 per Delivery Unit Best for Enterprise programs, multi-phase platforms
> — https://azumo.com/get-pricing · Rate cards, Platform column

**Seat model** · Partly — Services are sold by Delivery Unit or by engineer rather than by seat, and the only per-seat statement is Valkyrie's one price per developer, with no stated minimum or maximum anywhere.

> Predictable Pricing One price per developer
> — https://azumo.com/valkyrie · "Choose Your Open Weight Model" section, Predictable Pricing tile

**Pooled budget** · Partly — Valkyrie sells reserved concurrent agents shared by the team rather than per-user token balances, and work beyond the reservation queues instead of costing more; no per-seat balance is described.

> You buy concurrent agents, not tokens, so the bill is the same in a heavy month and a quiet one.
> — https://azumo.com/valkyrie · "How Valkyrie Works" section, "Valkyrie Runs it" step

**Admin controls** · Partly — Agent limits are set in the build rather than by an administrator in a console; Valkyrie mentions enterprise guardrails and key management but no page describes admin policy controls over what an agent may do.

> Valkyrie gives your team the power of open-weight models with enterprise-grade operations and guardrails so you can ship faster with confidence.
> — https://azumo.com/valkyrie · "Choose Your Open Weight Model" section, intro paragraph

**Audit log** · Partly — Agents they build log every action with reconstructable inputs and Valkyrie exposes per-request usage, but neither page describes an administrative audit trail of console or account actions.

> Full action logging with reconstructable inputs
> — https://azumo.com/artificial-intelligence · "Workflow Automation Across Systems" use case, "What you get" list

## What happens to your code

**Opt out of training** · Yes — Valkyrie states customer traffic never trains a model another customer can reach, prompts and completions are discarded at request end, and logging stays off until the customer enables it.

> Your traffic never trains a model another customer can reach. Improving your own model runs on your data, in your instance, and you approve each cycle.
> — https://azumo.com/valkyrie · "How your data is handled" section, "Your data trains nothing else"

**Getting out** · Partly — Fine-tuned weights are exportable and client code stays in the customer's own repositories with clean IP assignment, but nothing is published about Valkyrie balance expiry or data export on cancellation.

> Clean IP assignment, NDAs by default, SOC 2 certified.
> — https://azumo.com/legacy-modernization · "Why Azumo" section, "Your code, your repositories" tile

**Certifications** · Partly — SOC 2 is the one attestation Azumo claims, backed by a dedicated security page and an AICPA shield, and GDPR appears only as personal-data handling in the privacy policy; no ISO 27001 claim exists on the site.

> At Azumo we take the security of our customer data seriously. We work every day to ensure that our security is aligned with industry security standards.
> — https://azumo.com/resources/security · Security page, "Azumo is SOC 2 Compliant" section intro

## What changed

- 2026-10-20 — Capabilities · Free tier · Price · Terms

## Keeping it current

These facts are re-checked against Azumo's own pages and updated when they change. If you represent Azumo, sign in and update the listing from your dashboard at https://theagentsindex.com/dashboard. The operations this site accepts from an agent are listed at https://theagentsindex.com/.well-known/agents.json — that catalog needs no key to read, and each entry states its own auth.