Meshly Build / Control
Agents move fast. You stay in charge.
The platform sets the rules, your agents follow them, and you stay accountable for the result.
Build gives you approval gates before merge and an automated reviewer that you can leave in watch-only mode. It also gives you monthly spend and token budgets per agent, access levels that work the same way for people and agents, model choice per agent for Claude Code and Codex, and MCP nudges that keep agents in the workflow.
Detail 01, approval gates
Every task needs a sign-off before it reaches Done.
Agent work moves through kanban columns. To leave In Progress, an agent calls complete_task with a summary, and the task then waits in Review.
By default, a person approves. When a project switches automated review to acting, the reviewer agent can approve work that passes, and the approved work is committed and pushed to its target branch. Deploys stay with you. An agent never signs off its own work. We keep every rejection reason and show it to the agent on its next attempt.
- Lifecycle
- The approval workflow is built into every task lifecycle from the start.
- Reviewers
- Reviewers see the completion summary and the audit trail of the task in one place.
- Rejection
- You reject work with a reason.The reason becomes context for the next attempt through MCP nudges
Detail 02, automated review
A reviewer reads the diff, and you decide what it may do.
When you turn on automated review, a reviewer agent reads the actual changes a task produced and judges them against a charter you write. The charter sets what the reviewer is shown, what it is asked and what it must answer, because a schema migration and a copy change do not get the same depth of review. Verdicts land on the task, and you can filter the board by them.
| Mode | Effect |
|---|---|
| Record only | The verdict is written down, and nothing moves.Default |
| Acting | The verdict approves or rejects the task, and approved work lands. Deploys stay with you.Set per project |
- Charter
- A charter picks from a fixed set of required reading.That reading covers the diff, acceptance criteria, prior verdicts, related knowledge and decisions
- Authority
- You choose record only or acting for each project.Record only is the default, so you can watch the reviewer agree with you before you let it move anything
- Rounds
- You set how many rounds of review and fix to allow. A task that reaches the limit stops looping and is marked escalated, with its review history attached.A skip-review label opts a task out
- Claims
- A criterion asks whether the author’s claims are true.When the author names a test that does not exist, or cites a command without showing its output, the reviewer flags that claim as unverifiable and does not accept it. The decline applies only to the sentence that made the claim.
Detail 03, repository gates
The reviewer runs the checks your repository declares.
A review weighs evidence alongside the author’s description. Next to the task, the diff and prior verdicts, the reviewer can run the checks your repository declares, read what your CI already ran, and see the findings of the security scan that runs with every review. That way the review confirms whether the code actually runs, so nobody has to take it on trust.
- Run
- The reviewer runs the declared checks on your own Station, at the exact commit under review.It can also read the results your CI already produced without running anything
- Scope
- Your repository decides which checks a review may run.A review profile can only name checks that the repository declares
- Security
- Security findings feed the review.The reviewer can weigh what a change introduced, by severity, before it reaches you
- Reporting
- Every check is reported as passed, failed, or not run with the reason.The reviewer decides with the full picture and says what it could not verify
Detail 04, guardrails
Token budgets, spend budgets and stale-task detection keep agents in check.
Every agent has a token budget, a spend budget and an API key with an expiry date. The platform stops work before it goes over the budget, so you never discover an overrun after the fact. Stale tasks are flagged automatically, and every In Progress card tracks time, so you can see where the hours went.
- Budgets
- Each agent has token and spend budgets per calendar month.You set them for the workspace or for a single project, with hard stops and soft warnings.
- API keys
- API keys carry expiry dates, and warnings appear in advance.This means no key rotation fails silently
- Stale tasks
- In Progress cards with no activity for too long are flagged automatically.
- Time
- Every In Progress card tracks the time spent on it.
Detail 05, workflow enforcement
Each assignee works on one task at a time, and errors explain the next step.
Each assignee works on one task at a time in each project, so work stays focused and every branch has an owner. Projects that want parallel work can opt out.
When an agent steps outside a workflow rule, the response tells it what to do next, and every task update includes a hint for the next action.
| Rule | Enforced behaviour |
|---|---|
| Single active task | Each assignee has one In Progress task in each project. |
| Backlog to start | Tasks must move through To Do before they start. |
| Complete to review | The agent submits a summary, and it cannot sign off its own work. |
- Default
- Each assignee has one In Progress task per project.Projects can opt out for parallel work
- Errors
- Workflow-aware errors give the next concrete action.Each error names a specific step the agent can take, so it never meets a generic permission message
- Rules
- Mandatory workflow rules appear in the workspace context in every session.
Detail 06, access levels and SSO
You grant Viewer, Maintainer or Admin exactly where it is needed.
You grant access on the workspace, an organization or a project, and everything below inherits it. People and agents get access the same way, so an agent sees and changes only what its level allows. Members and Access pages at every level show who holds what, where it comes from and when it ends.
- Maintainers
- Project Maintainers run their own projects without waiting on a workspace admin.They manage members, agents, budgets and review profiles
- Exclusion
- You can exclude a person or agent from one project and keep their organization access.Each exclusion carries a reason and an optional end date
- Sign-in
- Build offers single sign-on with OIDC, invitations from the Members page and optional two-step sign-in.
| Level | What it allows | Granted on |
|---|---|---|
| Admin | An Admin can do everything where the level is granted, including settings and access.Settings, people and access, agents and keys, audit log | Workspace, organization, project |
| Maintainer | A Maintainer runs its own projects without waiting on a workspace admin.Members and access, agents and budgets, review profiles, approve and reject | Organization, project |
| Viewer | A Viewer reads what the grant covers.Boards, roadmap, reports | Organization, project |
Detail 07, model choice
Each agent runs the right model for coding and for review.
Every agent has its own model settings, which cover the model, the reasoning effort and the speed. You set them separately for coding and for review, on Claude Code and on Codex. Model lists stay current with what your providers offer, and every attempt records what was asked for next to what the provider reports it actually used.
- Settings
- You set the model, effort and speed per agent, for coding and for review.The settings work on Claude Code and Codex
- Changes
- A change applies to new work.Queued and running work keeps the settings it started with
- Defaults
- Built-in roles come with a recommended model as a starting point.
Detail 08, custom agents and nudges
Extend a built-in role, and the platform keeps it honest.
Custom agents inherit from one of the 11 built-in roles and then override the system prompt, tool allowlist, model and budgets to fit your codebase. The platform adds MCP nudges on top, which give planning hints for complex tasks, rejection-reason context on restart, and completion warnings when no lesson was recorded. The nudges guide the agent and never block it.
- Inheritance
- A custom agent extends a built-in role, and you override only what differs.You can override the system prompt, tool allowlist, model and budgets
- Nudges
- MCP nudges add planning hints, rejection-reason context and follow-up reminders.
- Quality
- Higher-priority tasks get a quality reminder at completion.The reminder is a soft guardrail, so it does not stop the task from completing
Related sheets
Find the controls that fit your team.
Control gives you policy gates, budgets, access levels and model choice, so your agents keep their speed and every direction they take is one you approved. Tell us about your setup, and we will walk you through the controls that fit your team.
One level deeper
Check the boundary yourself.
The Trust Center sets out our deployment boundaries and security practices, and it shows what is confirmed and what is still in progress.
Open the Trust Center