Is this a catalog of integrations?

No. There is no library of pre-built connectors. When your agent needs something, Graft reads the service’s public documentation and writes a tool for that one job. Any service with public documentation is within reach, whether or not it has an MCP server of its own.

What does self-extending mean?

Your agent’s set of tools grows when a task needs something new and shrinks when a tool goes unused, with nothing installed by hand. Graft builds each tool on demand, tests it, and keeps it in your toolbox for next time. The idea comes from research on agents that improve their own setup; Graft applies it to one part, the tools, and only after each tool has passed a live test.

Will it work with my agent?

Any agent that speaks MCP, the open standard agents use to reach tools. Claude and ChatGPT add Graft as a connector and sign in. Hermes, OpenClaw and other agents take one server entry with a token from Graft’s console.

Where do my passwords and keys go?

To Graft, once, encrypted. Graft adds them to each call at the moment it is made. They never appear in a tool’s code, in the sandbox it runs in, in a chat, or in front of any AI model.

Does it repair tools automatically?

Not yet. Today Graft tests every tool against the real service before your agent uses it, and keeps fixing it until the test passes. Repairing a tool that breaks later, when a service changes, is the next thing we are building.

Yes. The core is open source under AGPL-3.0. The Docker Compose setup runs the same core as Graft Cloud, with your own AI provider key, and nothing leaves your machines.

Graft Cloud is free for a limited time during the alpha. Paid plans will be priced per tool built, never per call your agent makes. Self-hosting is free; you pay your own AI provider.

Backed by

Combinator

Your agent grows
the tools it needs.

Ask for something
it has no tool for.

Graft will build it, test it against the real
service, and hand it over. Graft makes the
agent you already run self-extending.

Paste the below into your agent

Help me set up Graft and get my first tool built…

Choose your agent to copy its setup guide. Paste it in to get started. Graft Cloud is free for a limited time during the alpha.

How it works

A tool for the task,
not a bundle for everything.

Your agent asks. Graft builds the smallest tool that does the job, proves it against the real service, and adds it to your agent’s list. Nothing else comes along.

01The task
Your agent needs
one more tool.
You ask about an invoice in Xero. Your agent has no tool for Xero, so it asks Graft for one.
02Connect
You connect Xero.
Your agent never
sees the key.
Your agent hands the sign-in to you. You connect Xero once, and the key stays with Graft: the tool, the sandbox it runs in and the AI that wrote it never receive it.
03Graft builds
Graft writes it
and tests it.
Graft reads Xero’s documentation, writes the smallest module that makes the call, and tests it against the real service without changing anything.
GraftSandbox
invoice-status / index.tsTS
1export default async (
2 input: Input, ctx: Context
3) => {
4 const res = await ctx.fetch(
5 "/api.xro/2.0/Invoices/" +
6 input.invoiceNumber
7 );
8 if (!res.ok) throw Error();
9 const r = await res.json();
10 return r.Invoices[0].Status;
11};
Checked
Proof read passed
Dry run passed
Published · xero__check-invoice-status
Tool ready. Back to ChatGPT.
Built and tested against the real service.
04Keep and reuse
Ready to use.
Yours to keep.
The tool lands in your agent’s list and your agent answers. It stays in your toolbox for next time.

Every tool, built.

Room to think

Fewer tools in the room.
More room to think.

Every tool your agent carries costs it attention on every message, used or not. Graft gives an agent the tools for the task at hand and takes back the ones it stops using.

Tools in your toolbox

Tools needed for this task

Every tool exposed

~200,000

tokens · 1,000 tool schemas

With Graft

~5,000

tokens · 5 task tools + Graft meta-tools

98%

Less context used by tools

195,000 fewer tokens in this example.

Illustrative estimates: 200 tokens per schema + 4,000 for Graft’s meta-tools. Squares compare tool context on the same scale. Actual usage varies; conversation and tool results are excluded.

Built for your agents

Works with the agent you already run

Graft connects over MCP and works with any agent that speaks it → Claude, ChatGPT, Hermes and OpenClaw among them. One entry in its tool list, and nothing else about your setup changes.

For people running agents

Keep your agent.
Grow what it can do.

Connect the agent you already use, (e.g. Claude, ChatGPT or Hermes). Choose which of your accounts it may reach, then let it ask Graft for the tools it needs.

Your agent → Graft → Your tools

Ask for what you need.
Connect your accounts.
Approve what changes data.

Claude
ChatGPT
Hermes
OpenClaw
Any MCP agent

For companies building agents

Give your customers’ agents
the same ability.

Graft’s core is open source. Build it into your product so each customer’s agent can acquire its own tools, with every person’s accounts and toolbox kept apart.

Person 01

Their agent
Their connections
Their toolbox

Person 02

Their agent
Their connections
Their toolbox

Explore the core

More reach. Clear boundaries.

Your agent goes further.
You keep control.

Your passwords stay
out of the code.
You enter a key or sign in once, and it stays with Graft. Graft adds it to each call at the last moment. The tool, the sandbox it runs in and the AI that wrote it never see it.
Anything that changes
data asks you first.
A tool that only reads runs freely. A tool that would change or delete something asks once, and your answer holds. You can make any tool ask every time.
One door
to the outside.
Tools run in a sandbox with a single route out: Graft’s proxy, which only reaches the addresses you approved when you connected the service.

Integrations

Agent access to every integration.

Connect your agent to every app you use, from Xero and Google Sheets to your internal tools. Tell your agent what you need; Graft builds the integration.

Explore integrations