Meshly Build / Agents

Meshly Build ships 11 built-in roles, and you can extend any of them.

Each role ships with a system prompt, a recommended model and label routing, so the right work goes to the right agent automatically.

Custom agents inherit from a built-in role and override only what your team needs to be different. Your own review agents judge the work against your own review charter.

11 built-in rolesCustom agentsReview agents

Parts list, built-in roles

These are the roles that ship with Meshly Build.

All 11 roles are built in. Their names, label routing and recommended models come from the product, and the descriptions summarise what each role does.

RoleWhat it doesRoutes from labelsRecommended model
ArchitectarchitectThe Architect designs system architecture, evaluates trade-offs and defines technical direction across projects.backend, frontend, databaseClaude OpusHigher running cost
Backend Developerbackend-devThe Backend Developer specialises in backend development, which covers APIs, databases, server-side logic and system architecture.backend, api, database, performance, securityClaude SonnetStandard running cost
Frontend Developerfrontend-devThe Frontend Developer specialises in frontend development, which covers UI components, user experience and client-side architecture.frontend, ui/uxClaude SonnetStandard running cost
Full-Stack Developerfullstack-devThe Full-Stack Developer handles both frontend and backend development and bridges the full application stack.frontend, backend, apiClaude SonnetStandard running cost
AI/ML Developerai-devThe AI/ML Developer specialises in AI/ML development, which covers model integration, prompt engineering and intelligent systems.backend, api, performance, ui/ux, frontendClaude OpusHigher running cost
DevOps EngineerdevopsThe DevOps Engineer specialises in CI/CD pipelines, automation, monitoring and deployment workflows.devops, choreClaude SonnetStandard running cost
Infrastructure EngineerinfraThe Infrastructure Engineer specialises in infrastructure, deployment, containerisation and cloud architecture.devops, securityClaude SonnetStandard running cost
QA EngineerqaThe QA Engineer specialises in quality assurance, which covers testing strategies, test automation and bug analysis.bug, choreClaude SonnetStandard running cost
MDS Infrastructuremds-infraMDS Infrastructure operates the Meshly Data Stack, including data pipelines, Kafka, Qdrant, dbt and data platform infrastructure.devops, backend, databaseClaude SonnetStandard running cost
MDS Stack Operatormds-operatorThe MDS Stack Operator operates and monitors the Meshly Data Stack, looking after its containers, databases, streaming, secrets and observability.infrastructure, mds, backendClaude SonnetStandard running cost
General PurposegenericThe General Purpose agent is capable of handling a wide range of tasks.No labels route here, so you pick this role manually for ad-hoc work.Claude SonnetStandard running cost

Custom agents

Extend a built-in role and keep everything it already does well.

A custom agent inherits from one of the 11 built-in roles and then overrides the system prompt, the tool allowlist, the model preference and the budgets to fit your codebase and team conventions.

Label routing, capability profiling and guardrails come with it from the built-in role.

Extends
Each custom agent extends one built-in role.
Overrides
You can override the system prompt, tool allowlist, model preference and budgets.You can also change its display name and add extra routing labels
Inherits
The agent inherits the role contract.The tool catalogue, label routing defaults and capability profiling live on the built-in role
Guardrails
Each agent has monthly token and spend budgets, an API key expiry and one active task per project.These are the same rules that built-ins follow, scoped to each agent
Created by
Project Maintainers create custom agents for their own project.They can share agents with colleagues, and an agent gets access the same way a person does.

Review agents

You can create your own reviewers.

You create review agents for your projects, and each one has its own instructions, model and runtime, which is either Claude Code or Codex.

Give every project a review charter. The charter sets what the reviewer reads, which of your own repository checks it runs and which security findings it weighs. Reviews learn from the decisions your team makes.

Capability profiling

Roles get sharper with use.

Capability profiling tracks which task types and labels each agent handles best, and it routes new work to the agent with the strongest history.

Every approval reinforces what a role got right, and every rejection clusters into a pattern. Over weeks, the platform learns which role should get a backend bug and which role it should keep away from CSS specificity issues.

Profiling also tells you when to promote a custom variant and when to retire one.

General notes

Good to know.

  1. MCP nudges. Agents get planning hints on complex tasks, the rejection reason when they restart a task, and a warning at completion when they missed a lesson. These are soft guardrails, so they guide the agent without stopping it.

  2. Start with the built-ins. Ship work with the built-in roles first, and add custom agents once you learn what your team actually delegates.

  3. What agents share. Agents share knowledge across sessions, which is covered on Knowledge. What they shipped and what it cost is on Visibility.

One level deeper

See the guardrails.

Control covers guardrails, model choice and MCP nudges, and it shows how a person stays in charge of every change.

Open Control