Replit Agent

Platform · Last verified 2026-09-30

TL;DR

Replit Agent is the AI agent in Replit's hosted app-building workspace. For multi-agent work it has a task system: Agent splits a request into tasks, and each runs as its own Agent in an isolated copy of the project until you review and apply it. This entry covers that feature. It suits builders shipping several features at once.

Key facts

Replit Agent key facts. Data as of 2026-09-30.
Type	Platform
Languages / SDKs	Not applicable (hosted product)
License	Proprietary
Pricing model	Hosted, paid
Orchestration pattern	Supervisor
Repository	No public repository (closed source)
Website	replit.com
Documentation	docs.replit.com
Last verified	2026-09-30

Key features

Background tasks run as separate threads in isolated copies of your project, in parallel with the main thread, and change your main version only after you review and apply them. (source)
Agent proposes a set of tasks, each with a plan and a definition of done, on a board with Drafts, Active, Ready and Done columns; you accept or revise the plan before anything starts. (source)
Each task gets its own Agent that inherits the project context; Agent detects dependencies between tasks and holds dependent tasks until the prerequisite is applied. (source)
When tasks change the same file, Agent resolves conflicts automatically while applying, and the docs recommend reviewing the result after each apply. (source)
Concurrency depends on the plan: Starter has no background tasks, Core runs 1 at a time, Pro up to 10, Enterprise up to 64; extra tasks queue. (source)
Agent can connect to MCP servers (a curated list or a custom server), and a Replit MCP server lets ChatGPT, Claude, Codex, Claude Code and other MCP clients create and update Replit Apps. (source)
Teammates can be invited to a project and each person can start new threads with Agent on the same project. (source)

Architecture and orchestration pattern

Pattern: Supervisor

Replit Agent runs inside Replit's hosted Workspace and Project Editor; the user chats with it, sees a live Preview, and publishes from the same place. The docs describe two places where Agent works: the main thread, used for direction and decisions, and background tasks, which are separate threads that execute independently.

Multi-agent behaviour is the task system. When a request is large, Agent can split it into discrete tasks with plans, dependencies and a definition of done. Once accepted, each task runs its own Agent in an isolated copy of the project, so the main version is untouched. Independent tasks run at once up to the plan's limit, dependent tasks wait, and finished tasks show a work log, test results and a live preview. Applying a task merges it into the main version, with Agent resolving conflicts.

Context and memory are carried by the project itself (files, instructions, Skills) and by the main thread's context, which each task inherits. The docs do not describe agents talking to one another; the person acts as the director who approves, reviews and applies each task. Routines add scheduled runs that need a per-run budget.

Human in the loop

Approval is built into the task flow: you accept or revise the proposed tasks, approve each task's plan (Start building), review each finished task's work log, tests and preview, and then Apply or Dismiss it; a running task can be cancelled, which discards its work. Plan mode lets you brainstorm and plan without editing code, and the docs recommend applying tasks one at a time rather than batch-approving. Replit tells you before a request moves from Free Mode into a mode that incurs usage costs.

Harnesses it can drive

Claude Code (evidence)
Codex (evidence)

Protocols

MCP, A2A and AG-UI support for Replit Agent. See the full matrix.
Protocol	Support	Note
MCP	Yes evidence checked 2026-09-30	Both directions: Agent connects to curated or custom MCP servers (this page), and Replit hosts an MCP server at mcp.replit.com (Streamable HTTP, OAuth) for MCP clients (platforms/mcp-server page).
A2A	Unknown checked 2026-09-30	No mention of A2A or Agent2Agent in docs.replit.com/llms.txt, the Features export or replit.com/llms.txt. Closed-source, so no repository to search.
AG-UI	Unknown checked 2026-09-30	No mention of AG-UI in the same documentation exports, and Replit is not listed in the AG-UI protocol README.

Best for

Builders who want to develop several independent features of one app at the same time and review each before it lands in the main version. (shortlist)
Prototyping and shipping web or mobile apps from prompts inside one hosted environment, with Preview and one-click publishing. (shortlist)
Small teams that share a project and want each person to run their own Agent tasks against it.

Not for

Free Starter users who need parallelism: the docs say Starter does not support background tasks, and Core runs one at a time.
Work in your own repository and local toolchain; Agent works on Replit-hosted projects (import from other providers is possible).
Users who need agents that coordinate with each other; the docs describe person-approved, independent tasks.

Quickstart

Sign in at https://replit.com and start a project from a prompt; no local install is needed.

Install not yet verified by this site. What this means

# From the official docs (web app, no install)
1. Create an account or sign in at https://replit.com and describe an app in the prompt box.
   Free Mode is available on Starter, Core and Pro plans.
2. Test the result in Preview and improve it with a focused follow-up prompt.
3. Parallel work (Core: 1 task at a time, Pro: up to 10):
   - Expand the project in the sidebar and press + to create a task.
   - Describe one focused outcome, review the plan, then select Start building.
   - Open the board (Drafts, Active, Ready, Done); when a task is Ready, inspect
     the work log and Preview, then Apply (or Dismiss).
4. Optional: connect an MCP server under the Agent's MCP Servers settings.

Common pitfalls

Background tasks are not available on Starter; Core runs 1 active task at a time, Pro up to 10 and Enterprise up to 64; extra tasks queue.
Power Mode and Max Mode use monthly credits, and Routines run only in Power or Max Mode with a per-run budget; Economy Mode is gone (Power Mode replaced it at the same price).
Each task works on a copy, so applying one task changes the main version while others may still be based on an older one; review after each apply.
Two tasks that touch the same screen collide; the docs suggest doing overlapping work in sequence.
Cancelling a running task discards its work and cannot be undone.
Only connect MCP servers you trust.
Prices are subject to tax, and the docs list a cloud pricing update in August 2026.

Official quickstart

Pros

Isolated copies mean a running task cannot change your main version until you apply it. (source)
Dependency detection queues dependent tasks until their prerequisite is applied, so you do not track the schedule by hand. (source)
Task board and review flow give a visible plan, work log, test results and Preview for each task before applying. (source)
Includes hosting, Preview and publishing in the same environment, so no separate infrastructure is needed to see a task's result. (source)
Works with the Model Context Protocol in both directions, including a hosted Replit MCP server for other agents. (source)

Cons

Parallelism is gated by plan: no background tasks on Starter, one at a time on Core, up to 10 on Pro. (source)
The pricing page says Replit Agent is probabilistic and may occasionally make mistakes. (source)
Conflicts between tasks that edit the same file are resolved by the agent, so the docs tell you to check the app after every apply and to avoid parallel work on the same screen. (source)
Paid modes consume credits and the best models cost extra; the docs point to a spend-management page and note pricing updates. (source)
The free Starter plan is limited to Lite build with daily Agent credits and one published app that goes down after 30 days. (source)

Alternatives

FAQ

Does Replit Agent support MCP?

Yes, in two ways. Agent can connect to curated or custom MCP servers, and Replit hosts an MCP server at mcp.replit.com so clients such as Codex, Claude Code, ChatGPT and Claude can create and update Replit Apps.

Can Replit Agent run tasks in parallel?

Yes, through background tasks. Each task runs in an isolated copy of the project and is applied only after review. The limit is 1 active task on Core, 10 on Pro and 64 on Enterprise; Starter has none.

What does it cost?

The pricing page (read 2026-09-30) lists Core at $20 per month ($18 per month billed annually), Pro at $100 per month ($90 billed annually), and custom Enterprise pricing; the docs describe a free Starter plan with daily Agent credits. Power and Max modes use monthly credits.

Do the parallel agents talk to each other?

The docs describe independent tasks that share the project context they start with, plus dependency scheduling by the main thread. They do not describe agent-to-agent messaging.

Sources

Unknown fields: A2A and AG-UI: searched docs.replit.com/llms.txt, the Features export and replit.com/llms.txt for A2A, Agent2Agent and AG-UI and found nothing; closed-source, so no repository search. Scope: only the task system and background tasks were read as the multi-agent function; the docs pages read do not mention sub-agents inside a single Agent run, so none are claimed. Pricing: Core and Pro prices were read on replit.com/pricing; the Starter plan price and Pro credit tiers were not shown as numbers on that page, and per-mode credit costs were not read. Orchestration pattern 'supervisor' is inferred from the main-thread-plus-background-tasks description, not from a docs label.

Something wrong or out of date, or do you maintain Replit Agent and want this page removed? Report a correction or request removal.