In this article
Check the server address and tool list first
I’m the founder and CTO of MachGTM. When connecting an assistant, I would first check the endpoint and the work its registered tools perform. The hosted Mach workspace connector uses the signed-in member’s access. The public starter kit runs as a local process.
Model Context Protocol lets an assistant use tools exposed by a server. Read each tool’s actual inputs and effects before treating a returned answer as research or a saved change.
Use the workspace connection guide for Mach records and research. The public starter kit supplies five inspectable helpers for shaping inputs into example outputs. [1]
What the hosted workspace connector exposes
The connector address is https://machgtm.com/mcp. The connection guide shows how to connect an assistant and sign in to Mach. Settings, AI assistants lists your active connections.
Start with workspace_get to understand the workspace and available next actions. Brain tools read the seller's offerings, buyers and proof. Other tools inspect the search plan, leads, accounts, people, deals and today's tasks. The available list depends on the scopes granted to the connection and the member's current permissions.
| Question | Hosted workspace connector | Local starter kit |
|---|---|---|
| Connection | Remote HTTP endpoint at machgtm.com/mcp | Local stdio process |
| Account | Sign in to a Mach workspace | No Mach account connection in the handlers |
| Research | Can start workspace research after confirmation | Returns a checklist |
| Saved records | Uses authorised workspace routes | Returns text without workspace persistence |
| Cost controls | Confirmed plans and configured caps | No paid research calls in the inspected handlers |
An account_research_start call identifies a company already in Accounts and returns a plan with a maximum cost. After the person agrees, the confirmed call starts the run. Company research and people research are separate choices. The tool describes research as its task. Message writing happens later when requested or needed by a playbook.
A connection belongs to one member and one workspace. Check mach_status when the account or workspace is unclear. Revoking a connection removes its access. A fresh connection requires signing in again.
How a research request becomes authorised work
The hosted tools reuse the application's routes under the connected member's identity. Access is checked against the token scopes and workspace permissions. A connected assistant therefore needs permission for the particular record and action.
Guarded tools first return a plan. Confirmation must refer to that plan, on the same connection, with matching arguments. The implementation keeps plans for 10 minutes and applies configured spending caps. A repeated completed confirmation returns the saved result.
Read the plan's effect and maximum cost before confirming. Inspect the run result afterward. Starting a research job establishes that work was queued, while a completed job and a useful brief require their own checks.
These descriptions come from the implementation checked on 12 November 2026. Test your assistant host, connection scopes and workspace permissions using a known account before relying on the result.
The five tools in the free starter kit
The public package uses a local standard-input/output connection, commonly called stdio. Its entry point registers five tools and calls JavaScript handlers in the same process. The inspected handlers make no network requests and keep no Mach workspace state. [1]
| Tool | Input used | Returned output | Check before use |
|---|---|---|---|
| build_icp | Best customers and optional deal size | Largely fixed profile with copied inputs | Check the hard-coded market, geography and proof claims |
| research_prospect | Company, website, title and optional ICP | Research checklist and suggested questions | Obtain and verify the evidence separately |
| score_lead | Industry, headcount, technology and signals supplied | Rule-based score from 0 to 100 | Check the inputs and formula |
| draft_outreach | Name, company, signal, offer and optional proof | Templated message text | Remove unsupported fallback proof |
| build_sequence | ICP, offer and requested length | Slice of a fixed eight-touch schedule | Check length and timing before adapting |
build_icp returns a largely fixed sales-oriented profile. The code copies best_customers into based_on and places deal_size in its revenue field. company_description is accepted but unused by that handler. Treat the profile as a sample to inspect before applying it to a different market.
research_prospect returns a checklist of sources and questions. Supplying a company website places that value in the answer. It does not fetch the website or produce a sourced company brief.
draft_outreach interpolates supplied names, a buying signal, an offer and optional proof into templates. Read the resulting words carefully. If proof_point is omitted, the handler inserts a generic claim about having helped similar companies. Replace that line with verified proof or remove it before use.
Read the scoring formula before trusting the number
score_lead starts at 50 on a 0 to 100 scale. It adds 15 for an industry substring match in the supplied ICP, 10 for headcount from 10 to 500, 10 for a listed sales technology, and 5 per supplied buying signal. Headcount above 1,000 subtracts 10. The final score is clamped to 0 through 100. [1]
At 70 the label becomes Prioritise, at 50 Include in sequence, and below 50 Deprioritise. Those thresholds are rules in this helper. The implementation does not calibrate them against replies, meetings or purchases.
Its confidence label depends on the supplied signal count and headcount. A high label can therefore result from multiple unverified inputs. Check the source of each input before treating the score as evidence.
build_sequence slices a fixed schedule containing up to eight touches. It returns dates, channels and notes, with the supplied offer and ICP. It neither schedules those actions in a Mach account nor sends them. Check that the returned touch count and the requested length agree, because the handler does not constrain the requested length to the schedule's bounds.
Put it into practice
Test an output against known inputs
For a fictional local scoring example, supply industry SaaS, headcount fifty, an ICP containing SaaS, no sales technologies and no buying signals. The inspected formula starts at fifty, adds fifteen for industry and ten for headcount, producing seventy-five. This is a calculated example, with no tool invocation claimed. [1]
Two supplied signals would add ten and change the returned confidence label to High. That arithmetic leaves the signals’ truth unresolved. Check the evidence behind each input.
For the hosted connector, read one known account and inspect its research plan, target and maximum cost. Confirm when you intend to run that work. Then check the resulting status and brief against the source passages.
Check the result the selected tool can establish
The October operational reports describe Mach’s recorded research workflows. Their run costs supply no performance benchmark for the local helper functions. The account-research study states its actual contributing population.
The inspected hosted registration covers workspace work such as research, records and search planning. Check your connection’s current tool list for the action you need.
Mach’s playbooks have configurable message-review modes. Inspect the mode and permission for the intended workflow. The approval guide describes the layers.
Mach works with design partners and early adopters. Ask the team for current terms when evaluating paid workspace work.
Inspect the source and installation method
The checked starter-kit README documents npx -y github:verma6uc/mach-gtm-mcp, or cloning the repository, installing dependencies and running node index.js. Its local connection uses stdio. Follow a supported setup path for your assistant host. [1]
The hosted service uses the remote address https://machgtm.com/mcp. Follow the Mach connection guide and inspect the requested access during sign-in.
Write down the output you expect and compare it with what the tool returns. For a sourced brief, inspect its evidence. For a saved change, read the record back. For queued research, wait for the completed result.
Check the connection with one known account
Begin with a company whose identity and current record you already know. For the hosted connector, ask the assistant to report the connected workspace with mach_status, then read workspace_get and describe the available account-research action. Compare its answer with the workspace and tools shown in your connection.
| Task | Connection to inspect | Evidence to retain |
|---|---|---|
| Read your Mach workspace and company record | Hosted connector | Workspace identity, permissions and returned record |
| Plan company research | Hosted connector | Account identifier, plan and maximum cost |
| Inspect a research result | Hosted connector | Run status, brief and sources |
| Try an ICP or outreach text example | Local starter kit | Supplied inputs and returned template |
| Check a helper score | Local starter kit | Inputs, formula and calculated score |
Ask for the research plan for that account. Read the target, proposed work and maximum cost. Confirm the plan when you intend to start that work, then inspect the resulting status and brief. Record the connection, permission and observation date with your test.
For the local starter kit, supply a fictional company and predict the rule-based score from the formula above. Compare the returned text with the inputs you supplied. A checklist or interpolated draft should be assessed as that output.
Download the MCP connection checklist. It gives prompts and a place to record the expected output, actual output and any unresolved step. These are proposed checks, with no new live assistant test claimed.
How this post was made
Method: AI assisted this update after reviewing the public index.js and README plus the hosted research and confirmation paths on 12 November 2026. The source snapshot and code hashes are retained. The seventy-five-point example is calculated from the inspected formula. No fresh installation, assistant acceptance test or paid research call was performed.
Common questions
Which Mach MCP should I connect to for my workspace?
Use the remote address https://machgtm.com/mcp and sign in to Mach. The connection guide describes the setup. The free GitHub package is a separate local helper kit.
Does research_prospect search the web?
The local helper returns a checklist. Its handler makes no network request. The hosted account_research_start tool starts the workspace's research after a confirmed plan. [1]
Does the free kit send my inputs to Mach?
The inspected public handlers use their inputs locally and return text. They make no Mach API call. Your assistant host handles its own conversation data under that host's settings. [1]
Is the starter kit score a prediction of a sale?
It is a fixed arithmetic score from supplied industry, headcount, technology and signal inputs. The public implementation provides no measured calibration against purchases, replies or meetings. [1]
Does connecting an assistant authorise every action?
The hosted connector checks the connection's scopes and the member's current permissions. Guarded actions also require a matching confirmed plan. Inspect the tools available on your own connection.
How can I test the Mach MCP connection before relying on it?
Use one known account. Check the connected workspace, tool list, permissions and returned record first. Inspect the research plan and maximum cost before confirming a run, then read its status, brief and sources. The downloadable checklist records the expected and actual output.
Sources and further reading
- mach-gtm-mcp public source and installation instructionsVaibhav Verma / MachGTMBack to text



