Unflood for Claude
Build an agent in a Claude chat. Unflood tests it, versions it and, once a person approves, shares it with your team, from the same chat.
- Setup
- A custom connector, then Google or Microsoft sign-in
- Standard
- Remote MCP, OAuth 2.1
- Runs on
- Your Claude plan; Unflood never runs your agent
- Listing
- Custom connector
What your team gets
The agent stays where you built it. Unflood adds the part a team needs before relying on it.
-
From a chat to a tested agent
Describe the agent in Claude. Unflood records its instructions, tools and tests, runs the suites, and turns each failed case into an item to fix.
-
Nothing ships without a person
Each version is packaged and approved by a named person. If a release goes wrong, you return to the last good one.
-
Shared with the right people
Choose who can use, change or release each agent. Colleagues use it from their own Claude.
-
Taken to Claude Code
Export an agent as a Claude Code plugin. If someone edits the copy, Unflood flags the drift.
How it fits
Claude does the work. Unflood keeps the record and checks every change. They talk over one open standard.
Your Claude
- You, in a chat
- Claude's model runs the agent
- Your tools and data stay yours
Unflood
- Checks access and approvals on every call
- Keeps definitions, tests and evidence
- Versions, releases and the audit record
Claude Code plugin · export, with drift detection
Set it up
For you
Add the connector
In claude.ai, open Settings, then Connectors, and add a custom connector with Unflood's address.
Sign in
Continue with Google or Microsoft, then allow the access Claude asks for.
Ask Claude
For example: list my agents in Unflood. Claude calls query and shows the result.
For your company
An Owner adds Unflood once
In Claude's organisation settings, under Connectors. Adding it gives nobody access yet.
Set each tool's permission
Claude lets Owners mark a tool Always allow, Needs approval or Blocked. We suggest Needs approval for publish, approve, share and delete.
Members sign in
Each person connects with their own account, and sees only the agents shared with them.
Behind your MCP gateway
Route /mcp through it and paste the generated tool allowlist. The guide covers Cloudflare, Kong and others.
Exactly what Claude can do through Unflood
Every tool declares its side effect in its MCP annotations, so Claude and your gateway know which ones change things. The table below is generated from the same declarations, and a test fails if they drift apart.
26 tools: 1 read, 22 write, 3 remove or stop.
| Group | Tools | Side effect | Who decides |
|---|---|---|---|
| Read | query | Read | Claude, within what you may see |
| Build and test | create_agent, create_tool, create, update, evaluate, ask_agent, propose | Write | You, in the chat |
| Release | package, publish, approve, select (roll back), decide | Write | A named person. Unflood's own agents are refused |
| Share and access | share, delegate, connect_tool, refresh_grants | Write | You, for the people you choose |
| Run | run, raise, report, emit, resolve, configure | Write | You; runs stay under their limits |
| Remove or stop | delete, revoke_grant, cancel | Remove or stop | You; we suggest Needs approval |
What Unflood does not do
It does not run your agents
Claude's model runs them, on your Claude plan. Unflood holds their definitions, tests, evidence and releases.
It does not hold your working data
An agent's tools reach your own systems through a grant you consent to. The data stays there.
It does not read your chats
Unflood receives only what Claude sends in a tool call, nothing else from the conversation.
It does not approve anything
Its own agents propose and prepare. A person approves every release.
It does not keep your keys in an agent
Credentials are held as references, never written into an agent's definition.
For your security review
How the connection is secured, and who decides what.
- Protocol
- Remote MCP over HTTPS, one address: /mcp.
- Sign-in
- OAuth 2.1 with discovery (RFC 9728); each person signs in with Google or Microsoft and gets their own token.
- Token scope
- Each token is issued for Unflood's /mcp only (RFC 8707), and works behind your MCP gateway.
- Approvals
- Enforced by Unflood on every release, not by a prompt.
- Audit
- Who changed what, who approved it and why.
- Stopping an agent
- One command per agent version: it pauses triggers, revokes access and withdraws each host's copy.
- Certifications
- None claimed.
- Unflood: Agent made from this conversation 3m ago
- Unflood: Tests run on the new agent 2m ago
- Unflood: Released with your approval 1m ago
- Unflood: Shared with the finance team Just now
Questions
- Which Claude plans does it work with?
- Custom connectors work on every claude.ai plan, on the web, the desktop app and mobile. On Free you can add one custom connector.
- Is Unflood in Anthropic's connector directory?
- Not yet. Today you add it as a custom connector. The directory listing is prepared and waits for our release decision.
- Can my IT team keep Claude from changing things on its own?
- Yes. Every tool says whether it reads, writes or removes, so your Owners can set Needs approval or Blocked per tool, and your gateway can allow only the reads.
- What if we stop using Unflood?
- Your agents export as Claude Code plugins you keep. They run in your Claude, not in Unflood.
Try it on your own agent
Talk to us about a pilot for your team, built on one of your own agents.