BeliefState

BeliefState

[M8ven Score]

Bring cited investor theses, risks, catalysts and research changes into your AI through a hosted MCP server.

• Research with evidence: read attributed investment cases with source links, dates and explicit coverage gaps.
• Compare views: inspect differing theses, risks, conditions and publication-based outcomes.
• Follow changes: retrieve dated case revisions, engine assessments and research journals with their separate clocks.

Connection, authentication and pricing

Both research endpoints use Streamable HTTP. No local server or BeliefState SDK is required.

Account pay-as-you-go access is $20 for 500 nonexpiring credits ($0.04 per successful research request). One AI question can make several requests. Subscription allowances and existing grants have their own terms; check current account pricing. Connecting a client or creating a key does not add research credit.

Wallet pricing is separate: fourteen data tools cost 0.02 USDC each, compare_beliefs costs 0.50 USDC, and search_beliefs costs 1.50 USDC per call. Inspect the live catalog and exact quotes before authorizing payment. The wallet endpoint requires payment support; an ordinary OAuth-only client cannot pay its quotes.

get_tickers and search_companies are free on both endpoints. Account-only get_tracked_tickers is also free. Browser OAuth availability depends on the client's admission; check Connect AI before paying. The API-key setup below avoids assuming OAuth admission for a new client.

VS Code / GitHub Copilot · Tools · Wallet starter · Research workflow · Other integrations · Support · License

VS Code / GitHub Copilot

Use a current VS Code installation with AI features enabled and a Copilot agent session. This uses the current portable Copilot configuration format and a BeliefState API key; no local MCP runtime is needed.

1. Sign in to BeliefState account setup and obtain a developer API key for the account you want to use.
2. Make the key available as BELIEFSTATE_API_KEY in the environment of the host running the agent, using your private secret manager. For a remote workspace or container, provide it there. Start VS Code from that environment; setting it in an unrelated terminal after VS Code starts is insufficient.
3. Merge this configuration into your personal ~/.copilot/mcp-config.json (or mcp-config.json inside COPILOT_HOME when set). Keep any existing server entries.

{  "mcpServers": {    "beliefstate": {      "type": "http",      "url": "https://beliefstate.ai/mcp",      "headers": {        "Authorization": "Bearer ${BELIEFSTATE_API_KEY}"      }    }  }}

4. Start a new Copilot agent session. Open Chat: Open Customizations > MCP Servers to check beliefstate, and review any server-trust prompt. Enable its tools through Chat: Open Customizations > Tools if needed.
5. Ask “Which companies does BeliefState cover?” Expand the tool result and confirm BeliefState's free get_tickers returned coverage. Then ask a normal research question, such as “What's the investment case for Nebius?”, when you want to use research credits.

The free coverage call verifies discovery and authentication. A cited research answer additionally verifies research access and available allowance. A saved configuration or a plausible answer alone does not prove that a tool ran.

These steps follow the current VS Code setup guide and MCP configuration reference. The older VS Code servers/${input:...} format is different: interactive inputs are not forwarded to Copilot harness sessions. Never put a literal key in shared configuration, source control or chat. On October 29, 2026, a standard MCP client authenticated to the account endpoint with a Bearer key, listed all 19 tools and completed the free get_tickers call. The VS Code steps follow its current configuration reference; the native editor lifecycle has not been separately tested.

MCP tools

The account endpoint exposes the following 19 tools. The wallet endpoint exposes the same tools except get_tracked_tickers: 16 paid research tools and two free discovery tools. Live server cards provide current tool names and descriptions.

Use source dates, citations and coverage warnings from the returned records. Publication-based outcomes are not verified brokerage executions; missing coverage remains unknown.

Run a wallet-funded thesis audit

Use Node.js 22 or later on macOS or Linux. This example buys cited company research over HTTP using x402. It returns the research, source dates, coverage gaps and matching payment receipts for your own assessment. It does not execute trades or call a model. The default is one overview at 0.02 USDC; optional case opens cost another 0.02 each. An overview already includes selected cases, so open more detail only when you need it.

1. Clone this repository, enter it, and install the pinned dependencies:

   sh

   git clone https://github.com/Antcolombo/beliefstate-agent-integrations.gitcd beliefstate-agent-integrationsnpm ci --ignore-scriptsnpm test

2. Plan your requested company and spending cap without making network calls:

   sh

   node examples/wallet-thesis-audit.mjs --ticker NBIS --max-usdc 0.02

3. Check the live unsigned quote. This command cannot sign or pay:

   sh

   node examples/wallet-thesis-audit.mjs --ticker NBIS --max-usdc 0.02 --quote

4. Provide PAYMENT_WALLET_PRIVATE_KEY from your secret manager for a dedicated programmatic buyer wallet with native USDC on Base. Never export a recovery phrase or put keys in chat, source control, command arguments or shared logs. The buyer wallet is separate from BeliefState's receiving wallet.

5. Execute the approved purchase with a private state file:

   sh

   node examples/wallet-thesis-audit.mjs --ticker NBIS --max-usdc 0.02 \  --execute --state .private/nbis-wallet-audit.json

Execution checks free released coverage first and validates the quote's exact resource, receiving wallet, native USDC token, Base chain and price before signing. The full authorization is flushed to private state before dispatch. The state directory must be mode 700; its files must be mode 600. Keep the state private, including in backups. Research output and receipts also belong in your private workspace; payment signatures and keys are never printed.

To request detail for up to two case IDs returned by the overview, plan a separate workflow with --max-cases 2 --max-usdc 0.06 and a new state filename, then execute those same arguments. Choose this before starting; an existing state's ticker, cutoff, case limit and cap cannot change. The script reserves every signed purchase against the total cap, including unresolved attempts. It never uses illustrative discovery IDs or automatically buys model search/comparison. The cap covers service payments only; wallet funding, bridges and network fees are separate. Current quotes remain authoritative; changed terms stop execution.

Use --as-of 2026-10-27T12:00:00.000Z for a requested historical UTC cutoff. Otherwise execution fixes its start time as the cutoff and retains it for every case and retry. Dates, source qualifications, omissions and unknowns remain in the returned research. No result is a claim of complete coverage or performance.

If a request is pending, interrupted or fails receipt validation, keep the same state and rerun the same command. Recovery uses the saved exact method, URL, body and signature, without signing again; a replay-only run needs no private key. An expired authorization is still retained for recovery. Honor the saved retry_after/retry_at; the script prevents an early replay. Any unresolved purchase stops subsequent purchases, including on HTTP 402, 410 or 429. HTTP 410 alone does not prove a refund; inspect the private response and resolve the purchase with support before authorizing another one. Do not delete or replace uncertain payment state. A completed run returns its saved output without network calls. Exit status 1 means the run stopped or reached its cap; inspect its status.

An exclusive .lock prevents concurrent use of one state. After a hard process termination, confirm that no process still owns that exact state before removing only its .lock; preserve the payment state. Add --test for owner QA and retain it on every retry so those requests remain excluded from customer analytics.

Dependencies are pinned to @x402/[email protected], @x402/[email protected] and [email protected]. Offline tests use real SDK signatures with a public fixture key and a fake HTTP server function; they do not submit live payments. This starter does not add wallet signing to an MCP host; use an x402-capable host for wallet MCP.

Run a company research update

Requires Node.js 22 or later. No dependency installation is needed.

Get the examples and enter their directory:

git clone https://github.com/Antcolombo/beliefstate-agent-integrations.gitcd beliefstate-agent-integrations

Inspect a dated public research example without credentials or network calls:

npm run demo

The dated sample summarizes a verified October 27, 2026 NBIS run: one current case with a source link, seven journal events, and an intentional stop at the two-read ceiling. The pending continuation remains separate from a completed polling checkpoint. Coverage, omitted cases, unassessed fields and event domains remain explicit. The sample omits account identity, credentials and opaque tokens; it is a compact result summary, not a complete API response or current advice.

Plan the live workflow without making any requests:

node examples/research-changes.mjs --ticker NBIS --since 2026-10-20T00:00:00Z

Use the company and cutoff from your research question. The live workflow checks free released coverage, reads a compact current snapshot with up to three cases, then requests the ordered journal of research changes observed since that cutoff. The output retains attributed evidence, source links, event domains, clocks, quality gaps, history coverage, omissions and warnings. A journal event may be an engine assessment, distribution change or calculated outcome; inspect its domain before describing it as an author's revision.

For live account access, load your existing BELIEFSTATE_API_KEY from your secret store. Creating a key or connecting OAuth does not add research credit. Check current account pricing and available allowance before executing:

mkdir -m 700 .privatenode examples/research-changes.mjs --ticker NBIS --since 2026-10-20T00:00:00Z \  --execute --max-reads 2 --state .private/nbis-first.json > .private/nbis-first-output.json

Create .private with mkdir -m 700 .private before redirecting output into it. The script also creates private state directories when no output redirection is used. Keep returned research, state and checkpoints private. The key is neither saved nor printed. Use --test for owner diagnostics; ordinary customer usage should omit it.

--max-reads caps total account research dispatches in that state file, including an uncertain dispatch. At the October 27 account pay-as-you-go rate, two reads use $0.08 of credit; plan allowances and legacy policies can differ. This is a read-count ceiling, not a currency authorization. Wallet x402 data reads have their separate 0.02-USDC price. This script uses account REST and signs no wallet payments; the existing x402 quickstart owns that path.

If the journal has another page, the script saves the continuation and stops at the read ceiling. Resume the identical command with the same state file and a larger explicit ceiling, up to 20. Completed results are reused locally. The bounded current snapshot is not automatically paginated; its continuation and omission fields remain in the result.

After a completed journal window, start a later poll with a new state file:

node examples/research-changes.mjs --ticker NBIS --previous-state .private/nbis-first.json \  --execute --max-reads 2 --state .private/nbis-next.json

The old checkpoint stays unchanged until the new window is fully drained. resync_required, projection_catching_up and history_uninitialized need attention; they are not “no changes.” The example never creates a new request to recover an uncertain dispatch: account request IDs prevent another debit but completed response bodies are not retained for replay. Retain the state and ask support to reconcile its request ID before another attempt. A .lock file after a crash requires confirming the old process has stopped before removing it.

Run the offline safety checks with npm test.

Other integrations

Gemini is an optional configuration example, not the verified quick-start path. Check current client availability before authenticating or paying. The research skill supplies instructions; hosted access must be configured separately. The Postman collection contains no credentials or payment signer.

Explore free coverage and schemas without an account or wallet:

curl --fail-with-body https://beliefstate.ai/v1/researchcurl --fail-with-body 'https://beliefstate.ai/v1/tickers?q=NBIS'curl --fail-with-body https://beliefstate.ai/v1/research/get_belief

Support

Report integration issues in this repository. For account, service or private security reports, contact [email protected]. See the MCP setup guide for usage and beliefstate.ai for service information.

License

MIT covers the Gemini configuration, context and Postman collection. MIT also covers the workflow code. Research samples retain the service data terms and their stated provenance and limitations. The research skill uses MIT-0, as declared in its frontmatter. Hosted research access and data remain subject to BeliefState's service terms.