Conversational agents for messaging · built to survive production

The model proposes. Code decides.

Sanedryn is where agencies and builders build, test and run conversational agents on WhatsApp, Messenger, Instagram and the web. Every agent is a versioned spec. Every rule runs as code and is re-checked right before every send.

Your email app is opening. If it didn't, write to hello@sanedryn.com.

Now onboarding first agencies · For builders and businesses

An illustrative simulation: a message arrives, the model proposes a reply or a tool call, and code-enforced policy checks decide whether it is sent, held for approval, or stopped.
Why demos die

Your demo works. Production is where it breaks.

Today a messaging agent means gluing together six or more services. The demo holds. Then real customers arrive, and it breaks in the same places every time: human takeover, silent webhook death, the model skipping tools or inventing answers, message bursts, Meta onboarding.

Break 01

A person steps in. The bot keeps talking.

The owner answers from their phone, and the agent replies over them. The customer gets two voices.

The agent stops the moment a person replies from the inbox or the phone, and resumes only on release.

Break 02

Webhooks die quietly

An event goes missing or arrives twice. Nobody notices until a customer complains.

A durable queue, deduped webhooks and a send ledger, with a trace for every turn, so you can see what happened and why.

Break 03

The model improvises

It skips the tool, invents an answer, or does something the prompt asked it not to.

Tools fail closed and policies run in code. The agent can only use what a rule allows, and every send is checked first.

Break 04

Customers type in bursts

"hi" · "quick q" · "about my booking". Three messages, three half-answers.

Bursts of short messages are coalesced into one turn, so the agent answers the whole thought once.

How it works

Template. Test. Publish.

Every agent is a reusable template. A business's specifics go in as setup answers, so the agent you build for one client is ready for the next.

Start from a template

Pick a template and answer its setup questions. Goals, tone, language and links become part of the spec.

Feedback interviewTravel agentLead capture

Assemble and test

Layer blocks for personality, tone, rules, knowledge and tools. Connect any MCP server or HTTP API. Run conversations and scenarios in the simulator against the draft.

Publish and run

Publish a version to WhatsApp, Messenger, Instagram or a web chat widget. Watch per-turn traces. If something's off, roll back in one click.

The spec

An agent is a spec. Specs have versions.

Personality, tone, rules, knowledge and tools are layered, reusable blocks. Edit a draft, see the diff, test it, publish. One click takes you back to the last good version.

Claude · builder MCPIllustrative
>Make the travel agent's tone warmer, run the scenarios on the draft, then publish.
✓ draft updated · tone block
✓ simulator · scenarios run against the draft
✓ published

Build from Claude: a builder MCP server lets you edit, test and publish agents from claude.ai or Claude Code.

The platform

Everything between the model and the send button.

SYS·01 · TOOLS

Tools that fail closed

Connect any MCP server or HTTP API, plus a knowledge base with hybrid search. No rule, no tool. No wildcards.

SYS·02 · POLICY

Policies in code

Quiet hours, opt-out and action levels are enforced in code, not by the prompt, and re-checked right before every send.

SYS·03 · APPROVALS

Asks you first

Set an action level per tool. Sensitive actions wait for your approval, and can be undone.

SYS·04 · HANDOFF

Human takeover

When a person replies from the inbox or phone, the agent pauses. It resumes when you release the conversation.

SYS·05 · RELIABILITY

A runtime, not a flow

Durable queue, an exactly-once-style send ledger, deduped webhooks, burst coalescing and a trace for every turn.

SYS·06 · CHANNELS

Where customers already are

WhatsApp (official Business API), Messenger, Instagram, and a web chat widget or embeddable page. Messaging window rules respected.

SYS·07 · SIMULATOR

Test before live

Run conversations and saved scenarios against a draft before a single customer sees it.

SYS·08 · CONSOLE

Built for agencies

Many businesses in one console, scoped access for your team and your clients, statistics and traces.

SYS·09 · LANGUAGE

English and Hebrew

Agents and the console work in English and Hebrew, right-to-left ready.

Early access · now onboarding first agencies

No logos here yet. On purpose.

We're onboarding our first agencies one by one. Their stories will go here once they're real.

Early access

Build agents that
survive production.

Tell us where to reach you. We're opening Sanedryn to agencies and builders first, then to businesses.

Your email app is opening. If it didn't, write to hello@sanedryn.com.

Or write to hello@sanedryn.com