refund-policy.pdfp. 4
Customers may return any item within 30 days of the delivery date.
AI agent platform
Build AI agents that reach only what you allow
Every agent's access to knowledge bases, tools and the internet is set by you.
Build customer support agent, sales agent, research agent, internal knowledge and other business workflows.
Grounded answers
Every answer cites its sources in your approved knowledge bases
Upload your documents and the agent answers from them, quoting the exact passage behind each sentence. One answer can draw on several files at once. Hover a sentence to follow it back to its sources.
Answer
Sources
shipping-terms.md§ Delivery
An order counts as delivered on the day the carrier marks the parcel as delivered.
returns-checklist.md§ Condition
Accept the return only when the item is unworn and still carries its original tags.
How it works
Build an AI agent in three steps
Go from your first upload to a live agent API in three steps, all from the dashboard. If your agent does not need documents, skip step one and start with the model.
- 01
Upload documents (Optional)
Add policies, manuals, contracts or tickets. Commt indexes them into a knowledge base your agents can search. PDF, Word, Markdown and plain text are read.
- 02
Build the agent
Name it, pick a model, write the instructions in plain language, and attach the knowledge bases it should read. No prompt engineering background, and no code unless you want it.
- 03
Deploy with a limit
Test in the playground, promote from dev to production, then call it from your product with an API key. Set how many requests it gets before a customer ever sees it.
Connections
Connect the tools your agents need
Give your agents access to the business tools and services they need, and each agent reaches only the ones you allow. Connect through Commt's built-in integrations or extend an agent with MCP.
Twenty built-in integrations by category
Messaging and email
Messaging and email: Slack, Microsoft Teams, Telegram, WhatsApp, Instagram, Facebook, Reddit, Gmail and Outlook connect an agent to where people already write, from team chat to a customer's inbox.
Work and knowledge tools
Work and knowledge tools: Jira, Confluence, Google Docs, Google Drive, Notion, Airtable and GitHub connect an agent to where work gets planned, written, stored and tracked.
Customer and sales tools
Customer and sales tools: Salesforce, HubSpot and Zendesk connect an agent to where customer records, deals and support tickets live.
Or connect another way
Zapier connects an agent to tools beyond this list, and an API key connects it directly to your own product.
Need a custom tool? Connect MCP servers to extend what your agents can do. Commt verifies custom MCP servers before they are enabled and alerts you when its checks identify security issues. Learn how MCP works.
Use cases
What teams build with Commt
One platform, one plan. Attach documents when the job needs them, and leave them out when it does not.
Customer support agent
Answers from your help centre, policies and past tickets, and says so when the answer is not there.
Returns
- An answer with the passage behind it
- A handover when the documents fall short
- One request per answer, counted
Readshelp centrepoliciespast tickets
Point it at the help centre, the billing and shipping policies, and a year of resolved tickets. It replies in the customer's own words rather than quoting policy at them, and the tone and refusals are yours to set. Nothing is invented to fill a gap.
AskCan I switch to annual billing halfway through the month?
AnswerYes. The unused days are credited against the new term.
billing-policy.pdf · p. 2
Lead qualification
Reads inbound enquiries, scores them against your criteria and drafts the first reply.
Returns
- A score out of 100 per enquiry
- The criterion that decided it
- A drafted reply, sent by a person
Readsinbound enquiriesyour criteria
Write the criteria in plain language: company size, budget, timeline, the industries you do not serve. Every inbound form, email and demo request is read against them as it arrives, so the queue is already ordered when your team opens it. Nothing leaves the building on its own.
AskScore this demo request: 12-person agency, no budget stated.
Answer42 of 100. Under the seat floor, and budget is unstated.
qualification-criteria.md · § Fit
Research assistant
Reads long documents and reports back with the passage behind every claim.
Returns
- One answer across many files
- File, location and passage per claim
- A shortlist of what to read yourself
Readsreportsfilings
Attach a folder of reports, filings, papers or transcripts and ask a question that spans all of them at once. PDF, Word, Markdown and plain text are read. A reviewer can check the work instead of taking it on trust.
AskWhich suppliers raised prices more than once last year?
AnswerTwo of nine. Both increases fall in the second half.
supplier-review-q4.pdf · p. 11
Internal helpdesk
Answers the questions your team asks the same three people every week.
Returns
- The paragraph that answers the question
- Nothing outside the attached knowledge bases
- The repeat traffic off three people's desks
Readshandbooksinternal wikisprocess docs
Expenses, onboarding, VPN access, holiday policy: the questions repeat every week and the answers already exist in writing. Attach those documents and the agent takes the traffic. An agent can only retrieve from the knowledge bases it is bound to, so team-only material stays team-only.
AskHow long do I have to file an expense from a trip?
AnswerThirty days from the last day of travel.
expenses-handbook.md · § Filing
Operations monitoring
Watches a feed, decides what matters and raises it with the context attached.
Returns
- Only the items your rules flag
- The surrounding context, already attached
- A reason on record for the rest
Readsqueueslogsshared inboxes
Point it at a queue, a log stream or a shared inbox. Each new item is read as it lands, against rules you write and can change without redeploying anything. The person on call opens a short summary instead of a raw feed at three in the morning.
AskAnything in the overnight queue worth waking someone for?
AnswerOne. Payment webhook failing since 02:14, 340 retries.
alerting-rules.md · § Severity
Document review
Checks contracts and submissions against a checklist you define, and flags what fails.
Returns
- Pass or fail on every checklist item
- The clause behind each failure
- Review time spent on the exceptions
Readscontractssubmissionsyour checklist
Define the checklist once: liability caps, payment terms, data clauses, the certificates a submission must carry. Every contract, tender response and supplier form is then walked item by item, in the same order, at the same depth, however long the document is.
Asksupplier-agreement-v3.pdf against the vendor checklist
AnswerThree of twelve clauses fail. No liability cap in § 9.
vendor-checklist.md · § Liability
The dashboard
Model, instructions, knowledge and limits on one screen
Every edit starts as a new draft. Change one thing, publish a version, promote it when it behaves.
Modelclaude-sonnet-5200K context · Pro
InstructionsYou answer customer questions using only the attached documents. If the answer is not in them, say so and offer to open a ticket.
Knowledgesupport-handbook128 files · 4,912 passages
Monthly limit2,000 requestshard stop
Build it once, run it anywhere
- Versions, not overwrites. Every change creates a version. Promote the one that works from dev to prod, roll back the one that does not.
- Same agent, many projects. Define an agent once and deploy it wherever you need it, each with its own data and its own limit.
- Swap models without rewriting. Instructions and knowledge stay put. Change the model and compare the answers side by side.
- Test before anyone sees it. The playground runs the real agent against real data, counted like any other request so there are no surprises later.
Models
One integration, every model
Frontier models from Anthropic, OpenAI and Google, plus open-source models for everyday volume. Your plan decides which ones an agent can reach, and switching between them is a dropdown, not a migration.
| Model | Provider | Context | Available on | Good for |
|---|---|---|---|---|
| claude-fable-5 | Anthropic | 1M | Enterprise only | Long autonomous runs, end to end |
| gpt-5.6-sol | OpenAI | 1M | Enterprise only | Multi-step coding, agentic work |
| claude-sonnet-5 | Anthropic | 200K | Pro and above | Long documents, careful reasoning |
| gpt-5 | OpenAI | 128K | Pro and above | General purpose, tool use |
| gemini-2.5-flash | 1M | Entry and above | Long context at everyday volume | |
| llama-3.3-70b | Meta · open source | 128K | Entry and above | High volume, everyday tasks |
| qwen3.8-max | Alibaba | 1M | Entry and above | Multimodal reasoning, agentic work |
Entry runs the efficient and open-source models, Pro adds most of the frontier ones, Enterprise has no restriction. Your plan sets how many requests you get each month, and every answer counts as one request. Switching model takes one click and leaves your instructions and your data untouched.
Limits
Agents stop at your request limit
Your plan includes a monthly request allowance. Commt checks the limit before each request, so an unexpected loop cannot consume the rest of your allowance overnight.
- A ceiling per project. Split the allowance and give a tighter limit to the projects that worry you.
- Alerts at 50, 80 and 95 percent. Nobody discovers the ceiling by hitting it.
- A clear failure. Blocked requests return HTTP 402 with a machine-readable code, so your product can degrade gracefully instead of hanging.
Support project · requests this monthlimit 5,000
0
01,2502,5003,7505,000
Running · within limit
Security and data
What an agent can reach, and what it can't
The boundaries every agent runs inside, and exactly what each one covers.
No message content by default
By default, no message text is written to storage. What is kept is the metadata behind your usage report: request counts, latency, which agent answered. Message text starts being written only when you switch on a time-boxed debug window, or when a conversation is flagged as a failure.
Set data retention
Conversation history is kept for a period you choose, from nothing up to a year, never longer. Set it once for the organisation; a project can tighten that period but never extend it. Invoices are the exception: they have to be kept for years, so that setting doesn't apply to them.
Guardrails on what an agent may say
Set the topics an agent must refuse, the data it must never repeat back, and what it does when it is not confident. Checked on the way out, before the answer reaches your customer. Available on every plan.
It can only call what you allow
Web search and page browsing are on by default, and you can switch them off when you create the agent. Beyond that, it can only call what you attach: retrieval against its knowledge bases, and any connector you add, such as Slack or Jira. With web access off, those connectors still reach only the service you configured, never the open web at large.
Custom MCP servers are verified before use
A custom MCP server runs in an isolated environment and is checked for security issues before it can be used. If a vulnerability is found, you are warned and the server stays unavailable until it is addressed. If nothing is found, you are told it is ready to use.
Hosted in the EU
Application, database and document storage run in the EU on private networking. Your files sit in our storage, your vectors in an isolated database, neither on the public internet.
Pricing
Plans and pricing
Every plan includes a monthly number of requests. When the allowance is used, your agents stop. No credits, no bundles, no expiring tokens.
Entry
Startups and Small Teams. Your First Agent in Production.
$349$244/month
Up to 5,000 requests a month.
Request early access- Up to 5 projects + 10 Agents
- 2 knowledge bases, 300 MB of documents
- Unlimited official MCP server
- 1 custom MCP verification per month
- Open-source and efficient models
- Basic RAG on your documents
- Guardrails and safety policies
- API keys
- Organisation-wide spending cap
- Email support
ProMost teams start here
Growing Teams and Serious Workloads. Agents Your Business Depends On.
$799$559/month
Up to 16,000 requests a month.
Request early access- Everything in Entry, plus
- Adds most frontier models
- Up to 10 projects + 20 agents
- 5 knowledge bases, 1 GB of documents
- Monthly 3 custom MCP servers definition
- Advanced RAG with reranking
- Dev and production environments
- Agent versioning and promotion
- Per-project spending caps
- Budget alerts
- Priority support
Enterprise
Enterprise and Custom AI Solutions.
Custom
Request volume set by your contract, invoiced.
Request early access- Everything in Pro, plus
- Every model, including the largest
- Private server or on-premises deployment
- Self-learning AI agents
- Knowledge bases sized to your contract
- SSO and SAML
- Custom retention and a signed DPA
- Bring your own model keys
- Uptime SLA and named support
- HMAC request signing
- High limit storage and members
- Security review and onboarding
- Annual invoicing
A knowledge base is one set of documents your agents search. Give each agent its own, or point several agents at the same one.
Prices in USD, billed monthly, cancel any time. Included requests apply to the billing month and do not roll over. A payment method is required before uploading documents, because indexing costs us money the moment it runs.
Early access
Get in before the doors open
Commt is not open to everyone yet. This round has 25 seats available. Leave your email and we will send you an invitation when it is, along with 30% off the plan you pick, locked for 12 months from the day you join.
Doors open 28 September 2026, 09:00 UTC, with 22 of 25 seats left.
No newsletter, and we do not pass your address to anyone.
Questions
Questions teams ask before they build
What is a private AI agent?
A private AI agent is one that reaches only the knowledge bases, tools and connectors you explicitly bind to it, and nothing else. On Commt, you set that scope when you create the agent, choosing which knowledge bases it reads from and which tools and connectors it can call, and every request it answers stays inside that scope.
Can a Commt agent access the internet?
Yes, by default: a Commt agent can search the web and open pages on every plan, alongside any knowledge bases, tools and connectors you have bound to it. Switch that off when you create the agent to keep it a private AI agent that reaches only what you have explicitly scoped to it, such as Slack or Jira, each limited to the service it connects to, and nothing on the open web.
Which tools can an agent reach?
An agent reaches only the tools and connectors you scope to it, chosen from twenty available. Examples include messaging and email tools such as Slack and Gmail, work tools such as Jira and Notion, and customer tools such as Salesforce and Zendesk, plus Zapier for reaching further tools and a direct API key for calling the agent from your own product.
Is RAG required?
No, RAG is optional on Commt: an agent can run on its instructions alone, or you can bind it to one or more knowledge bases so it retrieves from your own documents. When you do attach a knowledge base, the agent reads only the knowledge bases bound to it, never anything else in your account.
What data does Commt store?
By default, Commt does not store your message content, only the metadata that produces your usage report: request counts, latency and which agent answered. Message text is written only when you switch on a time-boxed debug window or a conversation is flagged as a failure. You set your own retention period for what is kept, and deleting a record means either removing it outright or anonymising it.
Where is Commt hosted?
Commt is hosted in the EU. Application, database and document storage run in the EU on private networking, with your files in isolated object storage and your vectors in an isolated database, neither reachable from the public internet.
Which models can I use?
Commt agents can use frontier models from Anthropic, OpenAI and Google, plus open-source models such as Llama and Qwen, through one integration. Your plan sets which models an agent can reach, and switching a model is a dropdown change, not a migration, so its instructions and knowledge bases carry over unchanged.
What does it cost to run an agent?
Commt uses fixed monthly plans with a defined number of agent requests. Entry includes 5,000 requests and Pro includes 16,000. When the included requests are used, agents stop instead of generating additional usage charges.
What happens when an agent hits its request limit?
When an agent hits its plan's request limit, further requests stop immediately and return HTTP 402 with a machine-readable error code, so your application can show a fallback instead of hanging. Alerts go out at 50, 80 and 95 percent of the limit, the check runs before each request rather than after, and upgrading takes effect immediately from the dashboard.
Do unused requests roll over?
No, unused requests do not roll over: the requests included with your plan apply to that billing month only. We would rather say so plainly than let you discover it on an invoice, and if you consistently use far fewer than your plan includes, you are probably on the wrong plan.
Can we use our own model provider account?
On Enterprise, yes: you supply your own provider keys, Commt stores them encrypted and never exposes them again, and you keep your existing provider agreement. Commt still handles retrieval, versioning, limits and logs on top of your own keys.
Get started
Build your first agent this week
Build an agent on your own data, watch how it answers, and set its limit before it ever reaches a customer.