configure_pipeline
Stages, and what closing means. Your agent decides the shape of your funnel instead of asking you to click through a wizard.
The CRM your agent runs
Everything a person can do here, an agent can do.
Indraft was built for agents from the first table, configuration included, where elsewhere the agent was added last and it shows the moment yours hits "you will have to do that in the UI". Every value it writes carries who asserted it, so you can tell your agent's claim from your own correction, and undo either.
No dead ends
Setup is a tool call, not a screen your agent cannot reach
The work that other CRMs reserve for a human in an admin panel is part of the protocol here. An agent inspects the schema, shapes the pipeline, defines the object types your business actually runs on, then fills the whole thing in.
configure_object_type
If your business runs on Candidates or Campaigns rather than deals, your agent defines the type and it gets the same history and undo as everything else. Name the field that identifies one and the same record submitted twice is one record, not two. Your own objects.
bulk_upsert
Companies, contacts, opportunities, interactions and tasks, in volume, from the context your agent already has.
revert_change
The agent can undo its own work at the field, and merge has an unmerge. Getting it wrong is recoverable, not a restore from backup.
46 tools over the Model Context Protocol, and the same surface on the REST API. Four things stay with a person on purpose: authorizing an agent, adding a teammate, paying, and destroying something permanently. Everything about your customer data is available to both.
The result
A record you can hand to an agent and still trust
Letting something write at machine speed only works if you can see what it did. Every value carries the actor that set it, and an agent's summary is stored as a claim by that agent rather than as something Indraft knows.
That is what makes the rest safe to delegate. You are not taking the agent's word for it; you are looking at who said what, and when.
Northwind Ltd
northwind.comStage
Negotiation person · dana, 2 days ago
ARR
$48,000 agent · mcp, from the signed order form
Owner
Dana Whitfield person · dana, 3 weeks ago
Call summary
Claim agent · mcp asserted this. Indraft did not verify it.
How every record stores attribution. Two values set by an agent, two by a person, and a summary labelled as a claim rather than a fact.
For the people who do not run agents
The same records, on a screen somebody can work
Not everyone on your team will drive an agent, and they do not have to. Everything an agent writes lands in an application a person can filter, read, correct and add to directly, with the same rules and the same history either way.
There is more of the application on who it is for.
Safe to run unattended
Built for something that retries, and runs beside itself
An agent works while you sleep. It times out and tries again, runs in parallel with itself, and acts on state that went stale while it was thinking. A CRM built for a person typing turns that into duplicates and lost edits.
- It retried, and you still have one company
- Creates take a stable idempotency key, so a repeated call returns the original result instead of a second Northwind Ltd.
- It arrived late, and did not overwrite you
- An update can carry the version it expected. A write built on stale state is refused rather than allowed to bury a newer edit.
- It was unsure, and it asked
- Companies match on domain and contacts on email. Name similarity matches nothing. An ambiguous upsert returns the candidates and commits nothing.
- It got one wrong, and you undid it
- Field-level changes can be compensated from the activity view. What cannot be undone is labelled before you click, not after.
One record
Many writers. One thing that is true.
Agents over MCP, your own services over REST, people in the web application, and bulk operations all write to the same canonical record, and it stays one record rather than four versions of the truth.
Companies, contacts, opportunities, interactions, tasks, and pipelines mean the same thing in every workspace. Agents configure stages, typed custom fields, tags, and roles inside those boundaries, so an agent that has run one Indraft workspace can run yours without being taught anything new.
One shape, whoever is writing
Companies, contacts, opportunities, interactions, tasks, and pipelines mean the same thing in every workspace, so a value written by an agent and a value written by a person are the same kind of fact, told apart by who asserted them rather than by where they came from.
What Indraft refuses to do
Four things it will never do behind your back
- It makes no model calls. There is no assistant inside Indraft and no inference on any code path, including search. Your agent supplies the judgment. Indraft validates it, stores it, and shows it attributed to whoever asserted it.
- It never guesses an identity. No fuzzy matching, no embeddings, no model-assisted merge. Two records are the same thing because a normalized domain or email says so, or a human decided.
- It takes no raw custody. No email bodies, no call transcripts, no attachments. Normalized state, short summaries, source references, and bounded evidence snippets.
- It does not rewrite history. A revert is a new entry. The ledger records the actor, the operation, the before and after of every field, and where it came from.
Pricing
Your agents are not the meter
A free workspace that completes a real setup, then $29 a month for a small shop and $149 when there are 25 people to sign in. Connect as many agents and clients as you want on any plan. People and volume set the price; how hard you run them does not.
Start with an empty workspace and an agent
Sign up, paste the endpoint into Claude or any MCP client, and give it the starter instruction. What you are judging is whether the CRM it builds is one you would actually keep.
Free plan, no card, and the whole setup completes without paying.