Bloom MCP server · build blueprint

What is built, and what is not

Not ready to ship.

3 things stand between this server and anyone but its author using it: a decision, a test run, and a second person to sign in.

  1. 01Nobody else can reach it

    It runs on a local development worker. There is no address to give anyone, and no production installation has been verified.

    Decided 25 Sept: plan a guarded deploy. A staging plan goes to the account holder before anything goes live.

  2. 02Two people sharing it is unproven

    Two service users, two identities each, two businesses each — never verified against live Bloom. Fixture tests are not evidence for this.

    Decided 25 Sept: plan a live two-user test. The account holder supplies the second Bloom login.

  3. 03Everyone would hand over their Bloom password

    Bloom publishes no delegated sign-in we can use. The account holder has decided not to ask, and to keep this server's own login and approval pages. Each additional person types their Bloom email and password into this server.

    Decided 25 Sept: not contacting Bloom. Instead a second person's sign-in is tested separately against the deployed server, and a refused sign-in says plainly that it was refused without guessing at a cause.

Checklist progress

44.3%

263 of 594 checklist rows ticked · 331 still open

The bar counts checklist rows, which is the most representative figure this repository holds — it is not a readiness score. Until the first of the three is closed, nobody else can reach this server at all. The other two mean that even once they can, letting them would not yet be safe. The rows also do not close evenly: most of what remains needs a live run against a real Bloom account, a deployment decision, or an owner decision.

Every blueprint capability, scored against the evidence in the repository. Verified means a tracker row ticked against a named artefact. The coral dot means at least one operation in that capability was proven live — not the whole capability. Fully proven and partly proven are always shown separately.

Updated 1 October 2026103 top-level capabilities594 checklist rows174 registered MCP tools492 API catalogue entries244 operations fully proven8 partly proven

What this MCP covers

174
Registered MCP tools
Registered in mcp-server/src/mcp.ts. Registered is not the same as working.
492
Bloom API operations
The documented and observed Bloom API surface we are mapping.
594
Implementation checklist rows
Build, security, live-testing and release requirements, including sub-items.
155 read adapters160 write entries38 excluded139 with no adapter

Where the build stands

Top-level capabilities only.

Verified — evidence recorded Partly built Blocked upstream Owner-blocked Not started At least one operation proven against live Bloom data

Proven against live Bloom

Built is not proven. A journal proves the action that ran, not the whole operation.

244
Operations fully proven
Every action it carries ran against a real Bloom account, with a journal. Of 305 operations this server can attempt.
8
Partly proven
At least one action proven and at least one not. All are writes, and none is finished.
53
Fixture-tested only
Adapters that pass local tests and have never run against Bloom. Not counted as working.
53
Run journals behind it
Distinct live runs in mcp-server/docs/evidence, each naming the operations it ran.

How the headline breaks down

252 operations have at least one action proven. That figure is the two headline numbers added; it is not itself a measure of finished work.

148
Read capabilities proven
A read has one action, so proving the operation proves it.
121
Write actions proven
Distinct operation-and-action pairs with a journal behind them. Deliberately not added to the reads above: they are different units.
96
Write operations fully proven
Every action proven. 104 have at least one, so 8 are only partly proven.
1
Proven, action unknown
One journal entry whose action was not recorded. Its operation id covers both create and duplicate and nothing recorded separates them, so the entry is credited to neither. The operation itself is separately fully proven.

Registry v1 0b0effca01a4, schema v1 — the binding set these figures were derived from.

At least one action proven — 148 of 153 reads, 104 of 152 writes Fixture-tested only — built, never run against Bloom
19 writes unverifiable — no enabled read can confirm the effect 3 writes unaddressable — Bloom returns no identifier to send them to 5 severe writes gated, and they are not alike. Three cash-payment writes refuse any write without a matching one-time owner approval, which the owner grants on a server page for that exact action; the flow is fixture-tested and no approval has ever been minted and spent against a real Bloom write. Two integration writes have no gate built and refuse unconditionally, and updateIntegration is additionally blocked on a hosted OAuth redirect that no approval unblocks.

Sending a message is not among these. bloom_message_write is a controlled test harness: the recipient is a single address compiled into mcp-server/src/message-send-policy.ts and the schema refuses any other, the adapter has never run against Bloom, and available reads cannot prove acceptance or delivery. Nothing on this page claims the server can email a client.

By product area

Capability board

Each square is one capability. Dots mark live proof.

Show by

The “What users need” view follows the journey through the MCP, from connecting Bloom to shipping it. The “How we build it” view groups the same capabilities by engineering workstream.

—Select a squareEach square is one capability from the blueprint.

Every capability

Show by

Bloom API surface

155read adapters, scoped and identity-checked
160write entries with journalled recovery
38explicitly excluded for privilege, safety or contract reasons
139catalogued and searchable, no adapter yet — but only 16 are backlog; the rest are deliberate non-adapters with a recorded reason

Rows with a live run

Capabilities with at least one operation proven by a live run. The list is kept by hand in the generator, not derived from the tracker, and most of these rows are still in progress. Runs used synthetic records on the owner’s own businesses; most were deleted afterwards and confirmed gone, but not all — three remain stranded, a CDN derivative still answers after its source was deleted, and a bloomer-qa project tag remains in the account vocabulary — all recorded against ACCEPT-11. One of the three cannot even be checked: Bloom returns HTTP 500 for the archived-form filter in both directions, proven across two businesses, so whether that record still exists is unknowable from here. A live tag covers the run that was made, not every action the row names.

    Not started

    Capabilities the tracker records as not started: no adapter, no implementation. Everything not listed here is at least partly built, blocked upstream, or owner-blocked — and partly built is not the same as working.