Replit
by Replit
Build, update and publish Replit apps from your AI assistant. Anthropic lists 5 tools; Replit documents 8, including publish_app. Every build tool hands work to Replit Agent, which bills real credits.
Verified connector
Listed by Anthropic as a partner connector in its Connectors Directory.
Connection checked by Agentman on .
Anthropic states this reflects the level of review a connector received, not a security audit.
Connect Replit via MCP
https://replit-mcp.com/anthropic/mcpWorks in any MCP-compatible client. In Agentman Studio it is one click — no config file to edit.
Replit Tools & Capabilities (5)
Tool names from Anthropic's directory listing. This server requires sign-in, so we could not read tool descriptions or parameter schemas.
Limits
- The directory tool list disagrees with Replit's in both directions. Five names against eight, overlapping on four.
publish_app— the deploy tool — is documented but not listed, andimport-claude-design-from-urlis listed but not in Replit's tool table. - We could not verify tool behaviour ourselves. The endpoint returns 401 to an anonymous handshake, so no safety annotations were readable. Every read/write classification here is Replit's or ours by inference from its documentation.
- Prompts are unconfirmed. Anthropic's snapshot lists no prompt names, and the gated endpoint blocked our own
prompts/list, so we cannot say whether the server serves any. - The build tools are natural-language dispatchers.
create_app_from_promptandupdate_app_using_prompthand a sentence to Replit Agent. Their effective surface is whatever Replit Agent can do, and the documentation does not bound that. There is no diff to review before the change lands. - An MCP call can spend money. Agent work is billed by effort with one checkpoint per completed request, and Replit publishes no per-checkpoint price. Set a budget limit before connecting.
- How the paid-action confirmation behaves through MCP is undocumented. Replit describes the prompt for its own chat interface; the MCP documentation is silent on it.
- Usage data lags up to 30 minutes, so the dashboard will not stop a runaway loop in real time.
- Publishing is client-dependent. Through a direct MCP connection it takes a separate tool; on ChatGPT, Replit documents that apps are published automatically.
- A rollback does not restore your database by default, and never restores the production database — that needs a separate point-in-time restore, with a window of 7 days on Core and 28 days on Pro.
- Plan gates apply. Replit documents Connectors, full build and Free Mode as unavailable on the Starter tier, and Starter is limited to one published app.
- The connection reaches shared apps, not just yours. Replit states you can work with apps you can edit, including apps shared with you — so an agent's reach may extend to a collaborator's project.
- Two endpoint URLs are in circulation. Anthropic publishes
/anthropic/mcp; Replit's docs publish/server/mcp. Both are live; the documentation does not explain the difference.
Frequently asked questions
How many tools does the Replit MCP server actually have?
Replit's own documentation lists eight; Anthropic's directory snapshot lists five. Three names differ. The docs add `search_apps`, `list_apps`, `publish_app` and `get_publish_status`, while the directory carries `import-claude-design-from-url`, which the docs table omits. The gap matters because `publish_app` is a deploy button, and it is absent from the listing a reviewer reads before approving the connector.
Does an MCP call to Replit cost money?
It can, because the build tools run Replit Agent, and Agent work is billed. Replit prices paid Agent work by effort, bundling a completed request into one billable checkpoint. Core and Pro subscribers get a Free Mode allowance that resets every five hours with a weekly limit. Replit states it asks for confirmation before a paid action starts, but does not document how that prompt reaches an MCP client.
Does update_app_using_prompt change my live published app?
No, not by itself. Replit documents publishing as a separate five-stage pipeline that ends by switching your URL to the new version, and `publish_app` as the tool that starts it. An Agent update changes the project and its preview. On the ChatGPT surface, however, Replit documents that Agent generates and publishes apps automatically, so that boundary is client-dependent.
Can I roll back a change an AI agent made to my Replit app?
Yes, through checkpoints. Replit documents Agent as creating checkpoints automatically at development milestones, capturing project files, conversation context, environment configuration and database contents. Rollback restores a checkpoint with one click and you can roll forward again. Two limits: a rollback leaves your database untouched unless you opt in, and the production database is restored separately.
Why is the Replit MCP endpoint on replit-mcp.com and not replit.com?
Replit runs the MCP server on a dedicated host, separate from the main site and from the OAuth issuer at `replit.com/oidc`. The path segment is per-integration rather than per-client. We confirmed on 19 August 2026 that only two paths are real MCP endpoints — the directory's `/anthropic/mcp` and the documented generic `/server/mcp`. Every other path returns the site's HTML.
What does ask_question do on the Replit MCP server?
It queries one of your apps, not Replit's documentation. Replit documents `ask_question` as asking Agent about an app without changing it, taking the app id and your question as required inputs. It is the read-only member of the Agent-backed group: useful for orienting yourself in a project you inherited before any tool touches its code.
Which Replit workspace does an MCP connection reach?
One workspace, chosen during the OAuth sign-in. Replit's setup steps for both Codex and Claude Code say to pick the workspace to connect and then select Allow, and the Claude design handoff repeats the same choice. The connection reaches apps you can edit, which Replit states includes apps other people have shared with you, not only apps you own.
Sources
- Replit MCP Server documentation — https://docs.replit.com/platforms/mcp-server (fetched 2026-08-19 as Markdown via the
.mdsuffix; this is the URL Anthropic's directory publishes, and the source of the eight-tool table and the troubleshooting notes) · retrieved 2026-08-19 - Replit AI Billing — https://docs.replit.com/billing/ai-billing (fetched 2026-08-19; effort-based pricing, checkpoints as the billable unit, third-party API pass-through) · retrieved 2026-08-19
- Replit pricing and plans — https://docs.replit.com/help/pricing-and-plans (fetched 2026-08-19; credit-pack prices, plan gates) · retrieved 2026-08-19
- Replit Free Agent Mode — https://docs.replit.com/help/free-agent-mode (fetched 2026-08-19; five-hour allowance reset, weekly limit, paid-escalation confirmation) · retrieved 2026-08-19
- Replit managing your spend — https://docs.replit.com/billing/managing-spend (fetched 2026-08-19) · retrieved 2026-08-19
- Replit Checkpoints and Rollbacks — https://docs.replit.com/features/version-control/checkpoints-and-rollbacks (fetched 2026-08-19; what a checkpoint captures, database opt-in, production restore carve-out) · retrieved 2026-08-19
- Replit "After you publish" — https://docs.replit.com/features/publishing/after-you-publish (fetched 2026-08-19; the five publishing stages and the URL cutover) · retrieved 2026-08-19
- Replit in ChatGPT — https://docs.replit.com/features/platforms/chatgpt (fetched 2026-08-19; automatic publishing on that surface) · retrieved 2026-08-19
- Replit in Claude — https://docs.replit.com/features/platforms/claude (fetched 2026-08-19) · retrieved 2026-08-19
- Import Claude designs into Replit Design — https://docs.replit.com/design/import-claude-designs-into-replit-design (fetched 2026-08-19; the consent-screen permission labels and what the import carries) · retrieved 2026-08-19
- Replit Agent Modes — https://docs.replit.com/features/agent/agent-modes (fetched 2026-08-19) · retrieved 2026-08-19
- Replit docs machine index — https://docs.replit.com/llms.txt (fetched 2026-08-19, 406 entries).
docs.replit.com/robots.txtallows the docs paths and carriesContent-Signal: ai-train=yes, search=yes, ai-input=yes— synthesis is expressly permitted on all three axes (fetched 2026-08-19).replit-mcp.comserves norobots.txt; the path returns the site's HTML shell · retrieved 2026-08-19 - Live RFC 9728 protected-resource descriptor — https://replit-mcp.com/.well-known/oauth-protected-resource/anthropic/mcp (200) and https://replit-mcp.com/.well-known/oauth-protected-resource/server/mcp (200); the root form 404s and the path-append form returns the endpoint's own 401 (all fetched 2026-08-19) · retrieved 2026-08-19
- Live RFC 8414 authorization-server metadata — https://replit-mcp.com/.well-known/oauth-authorization-server (fetched 2026-08-19;
issuerishttps://replit.com/oidc) · retrieved 2026-08-19 - Live anonymous probe — 2026-08-19:
POSTto the endpoint returns HTTP 401 withwww-authenticate: Bearer resource_metadata="https://replit-mcp.com/.well-known/oauth-protected-resource/anthropic/mcp". Path probes of/mcp,/openai/mcp,/cursor/mcp,/claude/mcpand/vscode/mcpreturned HTTP 200 withtext/html, not MCP responses · retrieved 2026-08-19 - Replit support — mailto:support@replit.com · Privacy — https://replit.com/privacy-policy
- Anthropic connector directory — https://claude.ai/directory/b952bcd7-554a-460e-9c8a-a81074f9d430 (snapshot 2026-08-16) · retrieved 2026-08-16
Use in Agentman
Connect once and your agents call these tools on their own — on a schedule, in a workflow, with nobody at the keyboard.
Open in Agentman StudioServer Info
- Category
- Developer Tools
- Developer
- Replit
- Tools
- 5
- Domain
- replit-mcp.com
Using Claude Desktop or another MCP client? Setup docs — the connection URL above works anywhere.