This is the working register of every AI agent and MCP server we've built and operate — most of them inside our own business first. When we scope your build, this is what we're reusing from.
Agent projects are quoted as bespoke work because most vendors are starting from nothing. We aren't. The retrieval, the guardrails, the audit trail, the connection layer to your mailbox or CRM — those are solved, running, and already carrying load.
What remains is the thirty per cent nobody can pre-build: your systems, your rules, your exceptions. That's where the Sprint looks, and it's the only part you should be paying a premium for.
12 of 25 published
Mail is where most business context actually lives, and it's the system agents are usually given least safe access to. Mailman exposes a mailbox as a defined set of operations rather than a raw credential, so an agent can find the thread it needs and draft into it without ever holding send authority.
Runs behind almost every client-facing agent we deploy. Drafts land in the drafts folder; a human sends.
Email approvals go unread for days. WhatsApp approvals come back in minutes. For businesses across India, the Gulf and Latin America, this is the difference between an automation that flows and one that stalls waiting on a human.
Handles inbound routing to the right agent, structured approval prompts, and escalation to a person when a reply doesn't parse.
The writing was never the bottleneck — the research was. This agent does the six minutes of looking things up that precede every ninety-second reply, then hands your team a draft with each claim traced to the record behind it.
Ships in shadow mode first: it drafts silently against live tickets so you can measure accuracy before anyone relies on it.
Read the support use case →Reps don't skip the CRM out of laziness — admin is unpaid time on a commission plan. This agent extracts stage, next steps, objections and competitors from the recording, proposes the field changes for one-click confirmation, and leaves a recap in drafts.
It proposes; it never silently overwrites a field a person set by hand.
Read the sales use case →Runs nightly against the pipeline: deals with no activity, stages contradicted by the last conversation, close dates that have slipped three times without comment. Managers arrive at the forecast call already knowing where to push.
Deliberately reports on deals, not on people. No rep scorecards ship by default.
Onboarding fails in the gaps between departments, not inside them. The orchestrator holds the whole sequence, triggers each task in the owning system, and escalates the one that hasn't moved by day three — the laptop, usually.
Same machinery runs offboarding in reverse, which is the version your auditors will care about.
Every HR team loses hours a week to questions already answered in a document nobody can find. This agent answers from the current policy version, cites the clause, and refuses to guess when the handbook is silent — which is how it avoids inventing an entitlement you never offered.
Where the answer depends on the individual, it reads their record rather than generalising.
Matches invoice to purchase order to receipt, clears the clean ones, and for every exception writes a sentence saying what disagrees and by how much. Finance teams don't need a queue of flags; they need to know which flag is worth opening.
Never posts to the ledger. It prepares the decision; a person makes it.
The workhorse behind half this register. Extracts fields from messy real-world documents and attaches a confidence score to each one, so downstream automation can route low-confidence values to a human instead of propagating a bad read.
Handles multi-language documents and photographs of documents, which is what actually arrives.
Reads commit activity, board movement and open blockers, then produces a written update that says what actually shipped and what's at risk — rather than restating the ticket titles.
We run this on our own delivery teams. It's the reason we trusted it enough to offer it.
Standards documents don't get followed; templates do. Scaffold generates a new service with your logging, auth, CI and structure already correct, so consistency is the path of least resistance rather than a review comment.
Most valuable to platform teams tired of enforcing the same conventions in every pull request.
Transcription is solved and mostly worthless on its own — nobody reads transcripts. This pulls out what was decided, who owns it and by when, files each item against the right project, and leaves the transcript where it belongs: in the archive.
Built for teams that switch languages mid-sentence, because ours do.
No published entries in that category yet. The unpublished thirteen include several — ask us in a Sprint.
Some are client-specific and under NDA. Some are too tied to our own internal systems to describe usefully. We'll walk through the relevant ones in a Sprint.
It's a standard adapter that lets an agent read from and act inside an existing system. It exposes defined operations — search a mailbox, fetch an order, update a record — so the agent works through a controlled interface instead of an open credential.
It decides what to do: which records matter, what the reply should say, when to stop and ask a person. An agent without MCP servers has nothing to act on. MCP servers without an agent are just plumbing.
A working system is typically two or three MCP servers supplying the systems, one agent applying judgement across them, and a guardrail layer deciding what it may do unsupervised. That's the shape of nearly everything in this register.
Five days, no cost. We map one workflow, tell you which entries in this register apply, and build a narrow demo against your real data.