Rules your team wrote in Slack,
served as MCP tools

Synpese reads the public Slack channels you add it to, finds the rules your team has stated there, and serves each one to your agents as a typed MCP tool.

Get started How it works

northwind/RulesExample workspaceServing 32 tools

Rules31 rules

Search topics, rules and channelsAll sourcesMost recent

TopicRuleChannelConfidence

Refund policyget_refund_policyRefund in 30 days; after that, finance approves#cs-billing0.92

Deploy processget_deploy_process24h on staging first, never on a Friday#deploys0.95

Oncall rotationget_oncall_rotationWeekly, hands over Monday at 10:00#eng-oncall0.97

Incident escalationget_incident_escalationSev-1 pages on-call; status post in 15 min#eng-oncall0.88

Expense limitsget_expense_limitsUp to $500 without approval, then finance#finance0.90

Discount approvalget_discount_approvalOver 20% off needs the head of sales#sales-deals0.86

Customer data accessget_customer_data_accessProduction data only via the audit console#security0.84

Leave requestsget_leave_requestsTwo weeks' notice for more than three days#people-ops0.81

Vendor onboardingget_vendor_onboardingSecurity review before any new vendor#security0.79

Release notesget_release_notesEvery customer-facing change gets a note#deploys0.74

Travel bookingget_travel_bookingBook 14 days out, economy under 6 hours#finance0.77

EditorAgentvia mcp

Can we refund an annual plan bought 41 days ago?

get_refund_policy({ plan: "annual", days: 41 })

→ { allowed: true, approvers: ["finance"] }

#cs-billing · confidence 0.92

~/code/billing-apimain

Reads fromSlack only · read-only

Slackinvite-only

Linearplanned

Notionplanned

GitHubplanned

ServesMCP · bearer token

Claude Codeclaude mcp add

Cursor.cursor/mcp.json

Other clientsHTTP + header

How it works

From a Slack post
to a tool call.

What it reads

Public channels you add the app to, going back 90 days. Posts by people only; bot posts are dropped. It never asks for DMs, private channels or files, and it cannot post, edit or delete anything.

What it misses

Thread replies. Only the post that starts a thread is read, so a rule agreed in the replies never becomes a tool. Channels the app was not added to are skipped, and the run lists them.

What it keeps

Each rule as a topic, the rule text, its channel and a confidence score from 0 to 1. Messages are processed in memory and not stored. A rule scored under 0.4 is dropped.

Rule text

Rule text comes from whoever can post.

A rule is written from whatever someone posted in a connected channel, guests and outside members included, and it reaches your agents as a tool description. Synpese checks the posts and the rules for text written to steer an AI, and holds back what it flags. That lowers the odds. It does not solve prompt injection.

What the checks catch and miss

Injection test

gpt-4o-mini · run twice

Crafted attacks10

8 held back1 no rule1 served

Legitimate rules12

12 served, none held back

Ten attacks written for the test and nine ordinary threads. The attack that got through read as an expense policy, with no instruction to an AI in it, and a false policy stated plainly gets through too. This is Synpese’s own test, not a detection rate for attacks in the wild.

MCP

One tool per rule, behind a bearer token.

Your endpoint serves list_topics and one get_ tool for each rule, all marked read-only. A rule that depends on the case takes typed arguments and answers with a decision. Every answer carries the rule text, its channel, its confidence and when it last changed. A request without the token gets a 401.

Example workspacenorthwind

ToolChannelConfidenceAnswers with

get_refund_policy#cs-billing0.92Decision

get_deploy_process#deploys0.95Decision

get_oncall_rotation#eng-oncall0.97Rule text

get_expense_limits#finance0.90Decision

get_release_notes#deploys0.74Rule text

get_discount_approval#sales-deals0.86Decision

The tools answer from the rules Synpese found, and nothing else. Once a day a run reads the 90-day window again, sends only new or edited posts to the model, and records what it read, sent and stored.

Example workspacenorthwind

Threads read, and sent to the model

4183 sent(last night)

4503001500

Sep 21, 2026Oct 20, 2026

Confidence

31 served

0.83mean(floor 0.40)

630

0.401.00

Channels

Held back

Rules by channel

31+4(27 on Sep 21)

ChannelSep 21Sep 28Oct 5Oct 12Oct 19Now

#cs-billing889999

#eng-oncall677777

#finance445566

Sources

Slack

Next run, UTC

06:14

One channel first.
See what comes out.

Add the app to a channel where your team states its rules, run a sync, and read the topics it finds. Nothing can call your endpoint until you issue a token. Slack access is invite-only for now. Synpese is not in the Slack Marketplace, so Slack limits it to reading 15 messages a minute from each workspace. At that rate, 90 days of a busy workspace takes hours. Sign up and ask for access.

Get started Request access

Pricing

Free while it is early.
One person runs it.

There is no paid plan and no card to enter. Synpese is built and run by one person, not a company, so an email reaches the person who wrote the code.

Paid plans

None exists. If one is introduced, it is published and explained before anything is charged.

Nonetoday

Read the terms

Details

Nothing to pay today
No card on file
Never switched on for you
Moving to it is your choice

Early accessStart here

Slack only, and invite-only for now. Linear, Notion and GitHub are planned, not started.

$0during early access

Get started

Details

Slack, read-only
list_topics plus a get_ tool per rule
Re-read daily at your own minute in the 06:00 UTC hour
Delete everything from Settings

Who runs it

Parbhat Kapila builds and runs Synpese alone.

1person, not a company

Email Parbhat

Details

No team and no investors
No SLA or support desk
Data requests handled by hand
Email goes to the builder

Use cases

What it is for.Where it stops.

Example workspacenorthwind

Can we refund an annual plan after 41 days?

Slack

get_refund_policy0.92

Refund, annual plan, day 41just now · #cs-billing

Deploy to production on a Friday17 seconds ago · #deploys

Expense over $50040 seconds ago · #finance

Answer policy questions from the policy

A support or ops agent calls get_refund_policy with the plan and the days since purchase, and gets back a decision with the rule it came from. When the rule does not settle the case, the answer says so.

What a tool sends back Where each rule comes from What the checks miss

Example workspacenorthwind

Slack grantRead-only

Scopes

3 read scopesno write

Channels read

northwind · 7 channels

Daily re-read

06:14 UTC

Token at rest

AES-256-GCM

Run historyLast 50

ExtractionDaily

Scheduled

Window

90 days

Changed

9 posts

Channels

#cs-billing#deploys#eng-oncall#finance

Steps

Skip unchangedScreenExtractEmbed

ToolsServed

get_refund_policyget_deploy_process

Give coding agents your engineering rules

Rules posted in #deploys or #eng-oncall, like no deploys on Fridays or when on-call hands over, reach Claude Code or Cursor as tools the agent can check before it acts. A tool is information, not a permission system, so keep risky actions behind a confirmation.

Connect Claude Code or Cursor Keep actions behind a confirmation The three Slack scopes

Schedule

Daily run

Your own minute in the 06:00 UTC hour.

Read 90 days

Every public channel the app is in.

Skip unchanged posts

A post seen before is not sent again.

Extract and screen

New and edited posts go to the model.

Serve on the next call

No redeploy. The server rebuilds per request.

Pick up changes with a daily re-read

Once a day, at a fixed minute after 06:00 UTC, Synpese reads the 90-day window again. A post that changed goes back to the model, and the tool answers with the new text on the next call. Nothing reconciles two posts that disagree, and nothing reports what changed overnight.

What a run reads What happens on disconnect What is not built

Example: a refund rule

Your refund policy is in Slack. Your agent has never seen it.

An agent starts each session knowing nothing about how your company works, and fills the gap by guessing. Synpese reads the channels where your team states its rules and gives agents those rules as tools, each with the channel it came from.

Example workspacenorthwind

#cs-billing

Dana 10:14

Annual plans get a full refund within 30 days. After that it’s prorated, and finance has to approve it.

Sam THREAD REPLY 10:21

Monthly plans don’t get refunds. They can cancel any time.

Synpese reads Dana’s post. Sam’s reply sits in a thread, so it is never read, and the monthly-plan rule never becomes a tool.

Extraction

1 rule extracted

Refund policy

Annual plans are refunded in full within 30 days, then prorated with finance approval.
Confidence 0.92: how plainly it was stated, not whether it is true.
Takes two arguments, plan and days, because the rule turns on both.

Screen

No instruction to an AI found, so it is served. The screen is a list and a model’s judgment, and it misses attacks written as ordinary policy.

What it becomes

get_refund_policy

Takes plan and days, and returns a decision beside the rule text, the channel, the confidence and when it last changed.

The tool an agent calls for this rule is

ƒget_refund_policy()

It answers with structured output:

decision, in the rule’s own fields: allowed, requires_approval, approvers.
ambiguous, true when the rule does not settle the case.
rule and source, the text and #cs-billing.
confidence and lastUpdated, 0.92 and when a run last refreshed it.

Example:

Your agent calls:

get_refund_policy({ plan: "annual", days: 41 })

Synpese returns:

{ "decision": { "allowed": true, "requires_approval": true, "approvers": ["finance"] }, "ambiguous": false, "source": "#cs-billing", "confidence": 0.92 }

Follow one rule

A refund rule, from the post where someone stated it, through extraction, to the answer an agent gets back.

Rules your team wrote in Slack, served as MCP tools

From a Slack postto a tool call.

What it reads

What it misses

What it keeps

Rule text comes from whoever can post.

One tool per rule, behind a bearer token.

AAsskk wwhhaatt yyoouurr tteeaamm wwrroottee ddoowwnn

One channel first. See what comes out.

Free while it is early. One person runs it.

What it is for.Where it stops.

Answer policy questions from the policy

Give coding agents your engineering rules

Pick up changes with a daily re-read

Example: a refund rule

Your refund policy is in Slack. Your agent has never seen it.

Rules your team wrote in Slack,
served as MCP tools

From a Slack post
to a tool call.

One channel first.
See what comes out.

Free while it is early.
One person runs it.