Agentic Inbox

Host email in your Cloudflare account and use a built-in agent to search conversations, manage mail, and draft replies.

Stars
★ 8.6k
Last updated
6mo ago
License
Apache-2.0
Primary language
TypeScript

At a glance

How it runs
Self-hosted serviceWeb appMCP server
Works with
Platform-specificClaude Code
Cost
Free software; you pay for model usage
Setup effort
High · needs real infrastructure
You'll need
npmCloudflare accountCloudflare-managed domainCloudflare WorkersCloudflare Email RoutingCloudflare Email ServiceCloudflare Workers AICloudflare AccessCloudflare Durable ObjectsCloudflare R2Shell / CLINetwork accessMCP Server
Typical use
An individual with a Cloudflare-managed domain who wants a browser-based mailbox running inside their own Cloudflare account.
Not a fit if
  • Teams avoiding Cloudflare vendor lock-in
  • Organizations needing per-mailbox authorization
  • Users without a Cloudflare account and domain
Maintenance
No commits for 6+ monthslast commit 2026-04-17+7.9 ★/day over 7 days

What does this agent do, and when should you use it?

Agentic Inbox is a self-hosted email client that runs entirely on Cloudflare Workers, combining a React 19 web interface with an integrated email agent. Cloudflare Email Routing delivers inbound messages to a Hono Worker; each mailbox is stored in a separate SQLite-backed Durable Object, while attachments reside in R2. Its side-panel agent uses the Cloudflare Agents SDK's AIChatAgent, AI SDK v6, and Workers AI to read inboxes, search conversations, draft responses, and send messages. New mail can trigger an automatically prepared reply, but the user must explicitly confirm every send. Production deployment requires Cloudflare Access, and anyone admitted by the shared Access policy can reach every mailbox, including external tools connected through /mcp.

The React SPA calls APIs exposed by the Hono Worker and uses WebSocket connections to reach the EmailAgent Durable Object under /agents/*. Cloudflare Email Routing passes inbound messages to the Worker; MailboxDO stores mailbox data in SQLite and places attachments in R2. The client supports sending and receiving mail, rich-text composition, reply and forward threads, folders, search, and attachments. The AIChatAgent implementation exposes nine email tools for reading and searching mail, creating reply drafts, and sending messages; it runs the Workers AI model @cf/moonshotai/kimi-k2.5 and provides streaming Markdown, visible tool calls, and persistent chat history. Each mailbox can also have a custom system prompt. An MCP server at /mcp lets documented external clients such as Claude Code and Cursor operate on a mailbox by supplying mailboxId.

  1. An individual with a Cloudflare-managed domain who wants a browser-based mailbox running inside their own Cloudflare account.
  2. A user handling substantial correspondence who wants an agent to search prior conversations and prepare replies for review before sending.
  3. A small team managing several addresses on one domain that values separate Durable Object and SQLite storage for each mailbox.
  4. An operator who wants automatic first-draft responses to incoming mail without permitting unattended message delivery.
  5. A developer using Claude Code or Cursor who wants to read or operate application mailboxes through /mcp and a mailboxId.

Key setup facts

Pulled from the install and usage notes below. The project README has the current values.

Config files
wrangler.jsonc

How do you install or deploy this agent?

You need a Cloudflare account with a managed domain, plus Email Routing, Email Service, and Workers AI. Cloudflare Access is required for production or shared deployments. For the hosted flow, use the repository's Deploy to Cloudflare process and enter the receiving domain as DOMAINS. For source deployment, run npm install, set the domain in wrangler.jsonc, create storage with wrangler r2 bucket create agentic-inbox, and run npm run deploy; use npm run dev for local development. The deploy flow provisions R2, Durable Objects, and Workers AI, but post-deployment work remains: enable one-click Cloudflare Access for the Worker, save the displayed POLICY_AUD and TEAM_DOMAIN values as Worker secrets, create an Email Routing catch-all rule that forwards to the Worker, and configure the outbound send_email binding.

How do you use this agent?

After deployment and the Access, Email Routing, and Email Service configuration, open the application and create a mailbox for an address on the configured domain, such as [email protected]. Use the web client to receive and send mail, search conversations, organize folders, and handle attachments. Open the agent side panel when you want it to inspect the inbox, locate messages, or draft a response. Inbound mail may generate an automatic draft, but sending always requires explicit confirmation. You can set a system prompt per mailbox, and agent chat history persists. External clients may connect to /mcp and select a mailbox with mailboxId; access remains governed by the application's shared Cloudflare Access policy.

What are this agent's strengths and limitations?

Pros
  • Email delivery, compute, mailbox storage, attachment storage, and agent inference are integrated into Cloudflare services, with the deployment flow provisioning R2, Durable Objects, and Workers AI.
  • Every mailbox receives its own Durable Object and SQLite database, while R2 handles attachments.
  • Automatic replies stop at the draft stage and always require explicit confirmation before sending.
  • The product combines a full email client with nine agent email tools, persistent chat history, streaming responses, and visible tool calls.
  • The documented /mcp endpoint allows external tools including Claude Code and Cursor to operate on mailboxes by mailboxId.
Limitations
  • Core operation depends on Cloudflare Workers, Durable Objects, R2, Email Routing, Email Service, Workers AI, and Access, so this is not a portable deployment for another cloud without substantial adaptation.
  • One-click deployment is incomplete by itself; Access secrets, an Email Routing catch-all rule, and the send_email binding still require manual configuration.
  • There is no per-mailbox authorization. Every user admitted by the shared Cloudflare Access policy can access every mailbox, including through MCP.
  • The documented AI implementation is tied to the Workers AI model @cf/moonshotai/kimi-k2.5; no alternative model-provider adapter is described.
  • Adoption requires a Cloudflare account, a managed domain, and several enabled Cloudflare services before both inbound and outbound mail work.

How does this agent compare with similar options?

Key facts side by side with the most closely related agents.

Agent Source review Form / cost Stars Updated Language Full support on
Agentic Inbox This agent 48 · Major gaps Self-hosted serviceFree + model costs ★ 8.6k 6mo ago TypeScript Claude Code
Talorys 83 · Good — ★ 1.5k 1d ago TypeScript —
Kody Personal Assistant Platform 76 · Good CLIFree + model costs ★ 937 1d ago TypeScript —
PipesHub Enterprise AI Context Platform 65 · Some gaps Self-hosted serviceFree + model costs ★ 4.1k today Python —

How does FollowAgents rate this agent?

FollowAgents source review · FARS-2.1
Major gaps
48/ 100 5-point scale 2.4 / 5
Trust 12/29
Reliability 5/14
Adaptability 10/18
Convention 10/18
Effectiveness 7/13
Verifiability 4/8
Why each dimension lost points
Trust12 / 29 · 2.1/5

The code and README show that API and MCP access is available to users passing one shared Cloudflare Access policy, with access to every mailbox and no per-mailbox authorization; least privilege is therefore limited (1). Auto-replies are saved as drafts and require human confirmation, supporting a score of 2 for confirmation. The README describes the main email, attachment, Durable Object, R2, Workers AI, and MCP paths (2). The packet does not show dedicated protection for sensitive email or dependency vulnerability review, so those score 1. AI tools can read and move email and discard drafts, while confirmation is explicitly described for sending only; external-effect controls are limited (1). Deleting a mailbox removes only its settings file, with code noting that DO data and attachment blobs remain, so rollback and cleanup are incomplete (1). Apache licensing and copyright notices provide some attribution, but no clear individual maintainer or verified publisher identity is established (1).

Reliability5 / 14 · 1.8/5

The README, types, and code generally support mailbox management and AI drafting claims, but the system prompt says the agent can only draft and cannot send, which does not fully align with the README's nine tools and human-confirmed sending flow (1). The package manifest names the main platform and dependencies, but the packet provides no lockfile or availability evidence (1). Some input, rate-limit, and Access errors have concrete responses or troubleshooting guidance, but a consistent failure-handling approach is not shown (1).

Adaptability10 / 18 · 2.8/5

The README defines self-hosted email, AI assistance, and MCP use cases, and describes multiple mailboxes and configurable prompts (2). The shared Access boundary is stated, but descriptions of the agent's capabilities and limits conflict (1). Auto-drafting is tied to new email and the prompt explicitly directs the agent to invoke the drafting tool, making the trigger reasonably clear (2). Setup and local development cover Workers, R2, Email Routing, Workers AI, and Access (2).

Convention10 / 18 · 2.8/5

The README clearly organizes features, architecture, configuration, deployment, and troubleshooting (2). It documents installation, deployment, and required post-deploy steps (2). Project and feature naming is mostly consistent across the supplied files (2). Deployment guidance and feature descriptions are present, but broader examples and FAQ material are limited (1). The README states the shared-access boundary, and code comments disclose incomplete deletion cleanup (2). The LICENSE and package metadata both identify Apache-2.0 (3). No version history or changelog appears in the packet (0). Copyright ownership is visible, but maintenance ownership and an update path are not clear (1).

Effectiveness7 / 13 · 2.7/5

The agent provides tools to list, search, read threads, draft, and organize email; drafts are saved for review in the UI, making outputs usable for ordinary tasks (2). It adds email search and drafting to a conventional email client, giving a clear incremental benefit (2). It requires a Cloudflare account, domain, Email Routing, Email Service, Workers AI, and Access, while the costs and configuration burden of those services are not quantified (1).

Verifiability4 / 8 · 2.5/5

README claims about architecture, configuration, and features have corresponding evidence in the supplied code and package manifest (2). The README, code, types, and manifest provide some cross-checking, but the packet contains no test or deployment records (1). Documentation presents product behavior, access boundaries, and setup requirements as claims; source code also shows concrete implementation and a TODO. This static review does not present unseen evidence as verified (2).

Risks and how to mitigate them
  • The README says anyone passing the shared Cloudflare Access policy can access every mailbox, and MCP clients can operate on any mailbox by supplying a mailboxId. Check that this trust boundary fits the intended users before deployment.
  • Mailbox deletion removes only the settings file in R2; code comments say Durable Object data and attachment blobs are not deleted, so data may remain after deletion.
  • The system prompt says the agent cannot send mail directly, while the README describes sending after human confirmation and the supplied tool listing is incomplete. Verify the actual send permissions and consistency between prompt and tool implementation.
  • The settings accept agentSystemPrompt without validation, as a code comment notes; custom prompts can change agent behavior.
Evidence confidence: Low Reviewed Nov 07, 2026 Reviewed revision 48039bb6785a
See the full review method →
Maintain this project? Add the FARS badge to your README

The badge shows the current review score and verdict and links back to this page. It refreshes about once an hour after a re-review.

FARS

[![FARS review](https://followagents.com/badge/agentic-inbox.svg)](https://followagents.com/en/agents/agentic-inbox)
<a href="https://followagents.com/en/agents/agentic-inbox"><img src="https://followagents.com/badge/agentic-inbox.svg" alt="FARS review"></a>

FAQ

Can it run on a server without Cloudflare?
Not as documented. Receiving, compute, database storage, attachments, authentication, and inference depend respectively on Cloudflare Email Routing, Workers, Durable Objects, R2, Access, and Workers AI.
Will the agent send generated replies automatically?
No. Incoming messages can trigger automatic drafting, but every send requires explicit user confirmation.
Are permissions isolated per mailbox?
No. Storage is isolated per mailbox, but authorization is enforced by one shared Cloudflare Access policy. Anyone who passes it can access all mailboxes.
What should I check after an Invalid or expired Access token error?
The documented likely cause is an incorrect POLICY_AUD or TEAM_DOMAIN secret. Turn one-click Access off and on again, obtain the latest modal values, and reset both Worker secrets.
Can external agent tools connect to it?
Yes, through MCP. The application exposes /mcp, and the documentation names Claude Code and Cursor as clients that can operate on a mailbox using mailboxId. Those clients inherit the same all-mailbox Access boundary.
Is Agentic Inbox free?
The software is free, but Agentic Inbox calls an AI model, and you pay that model provider for usage (or run a local model).
Is Agentic Inbox open source?
Yes. The source code is on GitHub (cloudflare/agentic-inbox) under the Apache-2.0 licence.
Is Agentic Inbox still maintained?
The GitHub repository has had no commits since 2026-04-17, more than six months ago. That can mean the project is finished or abandoned; check its issues before relying on it.
Who makes Agentic Inbox?
Agentic Inbox is published on GitHub by the account cloudflare. FollowAgents has not verified which person or company is behind that account.
Was this page helpful?
View on GitHub ↗ Install ↓

Compare agents like this one

The same FARS review applied across the shortlist this agent qualifies for.