AI and agents
Bring your own AI. Plug in your agents.
Hilsen works with the model you already pay for and the agents you already use. An agent can read and draft; only a person sends what it wrote.
AI providers
Your key, any model from the provider’s live list
Claude
Anthropic
GPT
OpenAI
Gemini
Google
DeepSeek
DeepSeek
Kimi
Moonshot
Qwen
Alibaba
Any OpenAI-compatible endpoint
OpenRouter, Groq, Together, or a model you host
Where AI helps
Four jobs, each with a person in charge
Qualification
A role, a relevance and the reason for each scraped member, with a confidence. Your edits always win.Conversation context
A summary of each thread with a suggested temperature, state and next step. Suggestions, never overwrites.Drafting
Draft a campaign message or an Inbox reply with AI. You review it before anything is sent.Autopilot
Answers early replies from your Playbook, checked before every send, and hands off at Interested.
A model per task
Cheap where it’s bulk, strong where it talks
Set a workspace default and, if you like, a different model per task. Bulk qualification can run on a fast, inexpensive model while Autopilot uses your best one.
- Qualifying bios in bulk
- A fast, cheap model
- Summaries and next steps
- Your workspace default
- Drafting messages
- Your workspace default
- Autopilot replies
- Your strongest model
Keys
No key? Rules mode.
Without a model, every AI feature runs on rules with the same outputs, so nothing in Hilsen waits for an AI budget.
- Keys are encrypted and only their last four characters are ever shown.
- Owners and admins manage keys and models. Everyone can see which model runs.
- Usage is logged per task and model (tokens and speed), never the content.
- If a key stops working or runs out of credit, owners are told once and Hilsen falls back to rules.
MCP and REST
Agents draft. People send.
Connect Claude, Cursor or your own agent to Hilsen’s MCP server with a scoped key. The same keys work on the REST API, and signed webhooks (HMAC-SHA256) push events like a new reply or a booked call to your own stack.
Claude Code
claude mcp add --transport http hilsen \
https://hilsen.si/api/mcp \
--header "Authorization: Bearer havn_sk_…"No send tool, on purpose
Keys carry read, write or draft scopes, and tools are filtered by scope. None of them sends a Telegram message.
The ten MCP tools
- search_leads
- Search the CRM with the same filters as the CRM table
- get_lead
- A lead’s full record: profile, notes, follow-ups, conversations
- list_conversations
- Inbox conversations by view, with the latest message
- get_conversation
- One transcript, with its AI context and pending drafts
- summarize_conversation
- A summary and suggested next step, stored as a suggestion
- update_lead
- Stage, state, owner, tags, notes and follow-ups, all logged
- draft_messages
- Personal drafts that wait for a person to approve
- create_campaign_draft
- A draft campaign. Launching stays with a person
- add_leads_to_campaign
- Enrol leads in a draft or paused campaign
- get_campaign_performance
- Sends, replies, reply rate and calls booked
Start with one group link.
Create a workspace, load sample data and run a campaign on simulated accounts before you connect a real one. Free while in beta.