Can Claude change my cap table after I connect it?

Not on its own. Capable documents that read tools answer questions while write tools create a proposal, and nothing on the cap table changes until an authorized admin approves that proposal inside Capable. The proposal, the approver, and the time are recorded in the audit log.

What exactly can Claude see once the connector is live?

What your own Capable role can see. Admins reach the companies they administer, read-only members can ask questions but cannot propose changes, and employees, advisors, and investors reach only their own equity through the my_equity view. The connector does not widen your permissions.

How do I cut off access later?

Revoke the client from Account and security in Capable, which invalidates its refresh token immediately, or disconnect the connector in Claude under Customize, then Connectors. Access tokens expire within the hour either way.

Does the connector work on Team and Enterprise plans?

Custom connectors using remote MCP are available on Free, Pro, Max, Team, and Enterprise plans, with Free limited to one custom connector. On Team and Enterprise, an Owner or Primary Owner adds the connector for the organization first, and each member still authenticates individually unless the connector uses a shared credential.

Why does Claude report a different ownership percentage than I expect?

Usually the denominator. get_cap_table returns fully diluted ownership by stakeholder as of a date, so an as-of date, an option pool that has not been topped up, or an unconverted SAFE can all shift the number. Ask for the as-of date and the class totals before you act on the percentage.

Claude Cap Table Connector Setup: Connect Capable Safely

To connect Claude to a live cap table, add Capable's remote MCP server as a custom connector in Claude — Settings, then Connectors, then Add, then Custom, then Web — paste the server URL https://capable.so/api/mcp, complete the OAuth sign-in, and then verify three things before you trust a single number: which company the connector is scoped to, what denominator the ownership percentage is using, and that every write arrives as a proposal an authorized admin must approve inside Capable. That last point is the one founders get wrong most often, because a chat that can draft a grant looks a lot like a chat that can issue one.

Claude's own documentation covers the mechanics. Connectors let Claude access your apps and services, retrieve your data, and take actions, and Claude inherits each person's permissions from the connected service — if someone cannot reach a record in the source system, the connector cannot reach it either (Use connectors to extend Claude's capabilities). Custom connectors built on remote MCP follow the same settings path, and Claude reaches your server from Anthropic's cloud infrastructure rather than from your laptop, which is why a server behind a VPN or firewall will fail to connect (Get started with custom connectors using remote MCP).

What generic connector guides leave out

Most published walkthroughs do a fair job on discovery and scopes. They explain where the Connectors menu lives, that Claude reads a server's authentication metadata before asking you anything, that OAuth is granted per person, and that org owners can restrict which actions a connector may take. That material is accurate and worth reading.

The reader task they skip is the one that matters for equity: Capable-specific setup plus verification of company context, denominator, and approval behavior. A generic guide will tell you the connector is connected. It will not tell you that get_cap_table returns fully diluted ownership by stakeholder as of a date, that model_round and exit_waterfall are simulations rather than ledger changes, or that the six write tools in Capable produce proposals rather than mutations. Those distinctions are the difference between a useful finance assistant and a very confident one.

Setup decision table

Who you are	Where the connector is added	Who authenticates	What to verify first
Individual on Free, Pro, or Max	Your own Claude settings, under Customize then Connectors	You	That the OAuth screen names the right Capable account and company
Owner or Primary Owner on Team or Enterprise	Organization settings, then Connectors, added once for the org	Each member separately, unless a shared credential is used	That org-level tool permissions are set before members connect
Member on Team or Enterprise	Nothing to add; the connector appears in your list after the owner adds it	You, on first use	That you are signed in as yourself, not a shared service account
Employee, advisor, or investor	Same as any member	You, with the email your company has on file	That your view is limited to `my_equity`

On Team and Enterprise plans, an Owner or Primary Owner enables a connector for the organization before members can authenticate, and enabling it does not grant anyone access on its own. Owners can also restrict which actions a connector may take across the organization — for example, allowing reads while blocking writes — and individual users cannot override that.

Adding the Capable connector in Claude

Open Claude, choose Customize, then Connectors.
Click Add, hover over Custom, and choose Web.
Name it something your team will recognize, such as Capable, and paste the server URL https://capable.so/api/mcp.
Click Continue and review the authentication settings Claude detected. Capable uses OAuth 2.1 with dynamic client registration, so Claude registers itself and there is no client ID or secret to copy.
Finish adding the connector, then click Connect and sign in on Capable's own authorization screen.
In a chat, open the plus menu, hover over Connectors, and toggle Capable on for that conversation.

If you work in Claude Code, the documented equivalent is claude mcp add --transport http capable https://capable.so/api/mcp. Any compliant MCP client can connect the same way, because the server is a standard remote MCP endpoint over HTTP.

The verification checklist

Run this before you ask a real question. It takes about two minutes and it is the whole point of the exercise.

list_companies returns the company you meant, with the role you expected. If it returns several, note which one you are asking about.
get_cap_table returns an as-of date. Confirm the date is today, or the date you intended, rather than a stale default.
The fully diluted total matches the total you see in the Capable app for the same date. If it does not, stop and reconcile before asking anything else.
The option pool status is what you expect. An un-topped-up pool changes every percentage downstream.
get_409a returns a current fair market value and its validity window, so you know whether a grant dated today is covered.
list_proposals shows what is already waiting for approval, so you do not stack duplicate drafts.
A write attempt lands as a proposal, not a change. Ask for a draft grant and confirm it appears in Drafts or in the proposal list rather than on the ledger.
Your role behaves as documented: read-only members can ask but cannot propose; admins can propose; employees see only my_equity.

Read versus write matrix

Tool	Type	What it does	What it changes
`list_companies`	Read	Companies you administer, with role and plan, plus companies where you hold equity	Nothing
`get_cap_table`	Read	Fully diluted ownership by stakeholder as of any date, class totals, option pool status	Nothing
`list_stakeholders`	Read	Everyone on the register with relationship, status, and portal state	Nothing
`get_stakeholder`	Read	Every security one person holds, with vesting computed as of today	Nothing
`list_securities`	Read	Securities of one type with outstanding and vested amounts	Nothing
`get_vesting_schedule`	Read	The dated tranches behind a single certificate	Nothing
`get_409a`	Read	Current fair market value, validity window, and history	Nothing
`model_round`	Read	SAFE and note conversion, pool top-up, price per share, dilution by holder	Nothing — it is a model
`exit_waterfall`	Read	Proceeds by class and holder at an exit value	Nothing — it is a model
`recent_activity`	Read	Latest transactions	Nothing
`export_links`	Read	Signed-in download links for Excel, CSVs, OCF, certificates	Nothing
`my_equity`	Read	Your own grants and holdings at every company	Nothing
`search`, `fetch`	Read	Find a stakeholder or certificate, then open it	Nothing
`propose_create_stakeholder`	Write proposal	Add a person or entity to the register	Only after admin approval
`propose_draft_security`	Write proposal	Draft a grant or issuance into Drafts	Only after admin approval
`propose_send_for_signature`	Write proposal	Send a drafted agreement for e-signature	Only after admin approval
`propose_terminate_stakeholder`	Write proposal	Record a termination date and open the post-termination window	Only after admin approval
`propose_cancel_security`	Write proposal	Cancel an outstanding security, with a reason	Only after admin approval
`propose_record_valuation`	Write proposal	Record a 409A determination with effective dates	Only after admin approval
`list_proposals`	Read	What is waiting for approval, what was approved, what was rejected	Nothing

Two things follow from that table. First, the read side is where the value is for most finance work: ownership questions, vesting questions, 409A coverage, and what changed this week. Second, the write side is a drafting surface, not an execution surface. A proposal is a reviewable artifact with a diff of what will change and who signs, and the cap table moves at the moment a human approves it.

A short safe prompt sequence

Paste these in order in a fresh conversation. Each one is designed to fail loudly if something is misconfigured.

"List the companies I can reach through this connector, with my role on each."
"For the company I administer, show the fully diluted cap table as of today, including class totals and option pool status."
"What is the current 409A fair market value and how long does it stay valid?"
"If we sold for $120 million today, what does each holder take home? Show the waterfall assumptions you used."
"Draft a 40,000 option grant for a new hire starting next Monday on our standard four-year schedule. Do not send it — just create the draft and tell me what is now waiting for approval."

Step five is the test that matters. If the answer describes a draft sitting in Drafts or a proposal awaiting approval, the connector is behaving as documented. If it describes a grant that now exists, stop and check your org-level tool permissions before you continue.

Troubleshooting

The connection times out or never completes. Custom connectors are reached from Anthropic's cloud, not your machine, so a server on a private network or behind a firewall will fail even if you can reach it from your laptop. Allowlist Anthropic's published IP ranges for an outbound-only path, or confirm the endpoint is publicly reachable.

OAuth completes but the connector returns nothing. Check that you signed in with the Capable account that actually belongs to the company. A personal account with no company membership will return an empty list, which looks like a broken connector and is not.

Claude reports a company you did not expect. list_companies returns every company you administer plus companies where you hold equity. Name the company explicitly in your prompt rather than saying "my cap table."

A percentage disagrees with the app. Compare the as-of date first, then the option pool, then any unconverted SAFEs. model_round is a simulation and will not match the ledger until the round is actually recorded.

A write never appears. Look in list_proposals and in Drafts inside Capable. If the proposal is there and unapproved, that is the designed behavior. If it is not there at all, your role may be read-only, or an org owner may have blocked write tools for everyone.

Limits worth knowing

The connector acts only while it is answering. The published connector tool list describes request-and-response work in a conversation. For nightly exports or event-driven flows, use a separate automation platform or Capable's documented REST API.

Access is scoped to your role and revocable. Revoking a client in Capable invalidates its refresh token immediately, and access tokens expire within the hour. On Team and Enterprise, org-level action restrictions narrow what anyone can do and cannot be overridden from a personal account.

Read scope is not a security audit. Anthropic states plainly that custom connectors can reach services it has not verified, and that you should connect only servers you trust and review tool approval requests carefully. Capable's own security page describes the connector's OAuth 2.1 with PKCE, short-lived tokens, per-company scoping, and the rule that write tools create proposals until an authorized admin approves. Read those pages rather than assuming, and see how access is scoped for the current wording.

Plan availability is published at the plan level, not connector by connector. Capable lists the Claude and ChatGPT connector on both Startup and Growth, and it works in full during the trial. For the exact tool list your account exposes, the connected account's own settings are the only authoritative source. Use Capable's pricing page to compare current plans if you are deciding.

Evidence note: everything here about Capable's connector comes from Capable's public documentation pages, including the full connector and tool reference. No customer data, private workspace, or non-public information was used.