Skip to main content
Glama

@ripar/skills

Ripar's agent-interoperability layer.

Two protocols, one package:

  • an MCP server so Claude (or any MCP client) can drive Ripar — agent to tool

  • an A2A agent card so other agents can find Ripar without a human wiring anything up — agent to agent

Everything it reads comes off three registries that are live on Algorand TestNet. Nothing here signs anything: a write comes back as an unsigned transaction for a human or a wallet to approve.

IdentityRegistry     769444119    who an agent is
ReputationRegistry   769444120    what payments have been credited to it
ValidationRegistry   769444121    what work is open, how it was judged, and what is escrowed

You can check those on the explorer right now: https://testnet.explorer.perawallet.app/application/769444119

A job carries two money numbers and they are not the same thing. The budget is what the client says the work is worth. The escrow is what they have actually handed to the contract. A job showing budget 1.0 and escrow 0 is unfunded — and that is the single most useful thing an agent can know before bidding on it. Every job this package returns reports both.


Quick start

git clone <this repo> && cd ripar-skills
npm install
npm run build
npm test          # 192 tests, 13 of which hit the real chain

Try it without an MCP client at all:

printf '%s\n%s\n%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"probe","version":"0"}}}' \
  '{"jsonrpc":"2.0","method":"notifications/initialized"}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"ripar_get_reputation","arguments":{"agentId":1}}}' \
  | node dist/bin/mcp-stdio.js 2>/dev/null

That returns agent 1's real score, read out of box sc_ on app 769444120 while you wait.


Related MCP server: Agent Directory MCP Server

Connecting it to Claude

Open your Claude Desktop config:

  • macOS ~/Library/Application Support/Claude/claude_desktop_config.json

  • Windows %APPDATA%\Claude\claude_desktop_config.json

Create the file if it isn't there, and add this block. Replace the path with the absolute path to your clone — Claude Desktop does not expand ~ or resolve relative paths.

{
  "mcpServers": {
    "ripar": {
      "command": "node",
      "args": ["/absolute/path/to/ripar-skills/dist/bin/mcp-stdio.js"],
      "env": {
        "RIPAR_NETWORK": "testnet"
      }
    }
  }
}

If you already have other servers in mcpServers, add "ripar" alongside them rather than replacing the object.

Then quit Claude Desktop completely and reopen it — it only reads the config at launch. You should see the tools appear under the tools icon. Ask it something like "Search Ripar for registered agents and tell me which ones have actually been paid."

Once this is published to npm, the same block without a build step:

{
  "mcpServers": {
    "ripar": {
      "command": "npx",
      "args": ["-y", "@ripar/skills"]
    }
  }
}

If the server doesn't show up: Claude Desktop logs to ~/Library/Logs/Claude/mcp-server-ripar.log on macOS. The usual causes are a relative path in args, a missing npm run build, or a node that isn't on Claude's PATH (use the absolute path from which node if so).

Other MCP clients

Anything that speaks MCP over stdio works — the command is the same:

node /absolute/path/to/ripar-skills/dist/bin/mcp-stdio.js
# Claude Code
claude mcp add ripar -- node /absolute/path/to/ripar-skills/dist/bin/mcp-stdio.js

The tools

Tool

What it does

Reads or writes

ripar_search_agents

List or search the IdentityRegistry by domain, id, or address

read

ripar_get_agent

One agent record, optionally with its score and jobs

read

ripar_get_reputation

Settled payments, USDC volume, validator verdicts

read

ripar_list_jobs

Jobs on the ValidationRegistry — budget and escrow both

read

ripar_settlements

Real USDC transfers, next to the score the chain actually holds

read

ripar_quote_endpoint

Ask a paid endpoint what it costs, without paying

read

ripar_call_endpoint

Call an endpoint; returns the 402 challenge if it charges. checkPayee: true refuses to forward a payment to an unchecked payee

can spend

ripar_payee_check

Before paying an address: exists? can receive USDC? paid over x402? Bazaar-listed? $0.01, cached 24h

can spend

ripar_post_job

Compose a post_job call and return it unsigned

returns a transaction

ripar_fund_job

Compose the two-transaction group that puts money in escrow

returns a group

ripar_settle_escrow

Compose a release_escrow or refund_escrow call

returns a transaction

The four that aren't reads

ripar_post_job returns base64 msgpack plus a plain-language summary:

{
  "signed": false,
  "unsignedTxnBase64": "iqRhcGFhlMQE6xJnHsQiACBdanwFPa6OATBBTNfKO3sHnSiP…",
  "txId": "IM2SP7CUFPBNC5BSJJEKG2FDCLPXPSKZKY53IZHTOAVERKEOMMUA",
  "method": "post_job(byte[],uint64,uint64)uint64",
  "summary": "Open job #3 on ValidationRegistry 769444121 with a budget of 2.500000 USDC, committing to spec hash 5d6a7c…, to be judged by agent 1. Signing this makes UBB4… the job's client.",
  "boxes": ["jb_3"],
  "nextSteps": ["…", "Sign it with the wallet that holds `sender`; this package holds no key and cannot sign.", "…"]
}

ripar_fund_job returns two transactions, because fund_job is a group and only works as one:

{
  "signed": false,
  "groupId": "<base64 of the 32-byte group id BOTH transactions carry>",
  "method": "fund_job(axfer,uint64)uint64",
  "summary": "Move 1.000000 of asset 10458941 into escrow for job 3 … the money leaves KBDR… and is held by the contract at EJHY… until the work passes (release_escrow pays the assignee) or fails (refund_escrow returns it here).",
  "transactions": [
    { "index": 0, "kind": "axfer", "summary": "Transfer 1.000000 of asset 10458941 from KBDR… to the app account EJHY…" },
    { "index": 1, "kind": "appl",  "boxes": ["jb_3", "es_3"], "summary": "Call fund_job(axfer,uint64) … which reads the amount off transaction 0" }
  ],
  "nextSteps": ["Sign BOTH transactions … a group is invalid if any member is missing or moved.", "…"]
}

The contract reads the escrow amount off transaction 0, not from an argument — so the number it records is one the AVM has already validated. That is the same rule that stopped reputation being minted from bytes, and it is why this is a group rather than a call with a number in it.

ripar_settle_escrow composes the other end. release pays the assignee and is legal only on a passing verdict: the client may sign immediately, and anyone at all may sign once the dispute window has closed — because a validator who never comes back would otherwise freeze the worker's money for good. refund returns the escrow to the client on a failed verdict or a cancelled job, and anyone may sign that too, since the destination is read off the job rather than off the sender.

Both refuse, before a fee is spent, anything the chain would reject anyway: the wrong sender, the wrong status, or an escrow that is already 0.

ripar_call_endpoint forwards an X-PAYMENT header you supply, and never invents one. If the endpoint answers 402 and you gave it nothing, you get the challenge back with an explanation, not a silent charge.

Checking the payee first (opt-in, $0.01)

An agent that found an endpoint five minutes ago is about to sign a transfer to an address nobody has looked at. ripar_payee_check asks Ripar's paid endpoint — POST https://api.ripar.io/api/payee-check, $0.01 USDC on MainNet — whether it exists, can receive USDC, how often it has really been paid over x402, and whether it is Bazaar-listed. Facts from chain, no score.

It follows the same no-key split as everything else here:

// 1. no cached verdict, no paymentHeader → the $0.01 challenge to sign
{ "name": "ripar_payee_check", "arguments": { "address": "PBXELTAX…UINYAU" } }
// → { "source": "payment-required", "quote": { "price": { "amountDisplay": "0.010000 USDC", … } }, "note": "…" }

// 2. same call with the signed header → the verdict, cached for 24h
{ "name": "ripar_payee_check", "arguments": { "address": "PBXELTAX…UINYAU", "paymentHeader": "<signed>" } }
// → { "ok": true, "source": "check", "facts": ["Account exists.", …], "result": { … } }

// 3. now pay the endpoint, with the gate on
{ "name": "ripar_call_endpoint",
  "arguments": { "url": "https://…", "method": "POST", "paymentHeader": "<signed>", "checkPayee": true } }
  • Off by default. checkPayee only reads verdicts; it never pays for a check. With it on, the payment header is forwarded only if the address its signed transfer actually pays (decoded from the header, not taken from its unsigned payTo claim) has a passing verdict from the last 24h. Otherwise nothing is sent, and the error says which address to check.

  • Refuses when exists or canReceiveUsdc is false, or when requireListed is set and the payee is not Bazaar-listed. Everything else is reported, not judged.

  • Cached 24h per network+address in ~/.ripar/payee-cache.json — the same file @ripar/sdk's ripar call --check-payee writes, so a check paid from either clears the payee in both. RIPAR_PAYEE_CACHE moves it; a missing or corrupt file just means another check.

  • A check that fails (5xx, network, an answer that is not a verdict) is an error, and caches nothing — the payee stays unchecked, and the gate keeps refusing.

  • RIPAR_PAYEE_CHECK_URL overrides the endpoint. The payee-check endpoint is never gated on itself.

Why: an MCP server is driven by a model. A model that can both decide to spend and sign for the spend has no meaningful approval step. Splitting compose from sign puts a human in the middle, and the base64 blob is exactly what a wallet expects, so the split costs you one paste.

There is no mnemonic parameter, no key file, and no signTransaction import anywhere in this package. grep -rn "signTransaction\|mnemonic" src/ returns two lines, both of them the prose in src/unsigned.ts promising the code is not there.


Resources

The registries are also readable as MCP resources — documents a client can attach to a conversation without anyone deciding to call a tool:

URI

What it holds

ripar://registries

The three app ids, and the escrow terms read off the contract's own global state

ripar://agents

Every agent, with the URL its A2A card should be at

ripar://agent/{id}

One agent, its score, and the jobs it is on

ripar://jobs

Every job, budget and escrow side by side

ripar://job/{id}

One job, plus when its escrow becomes releasable by anyone

The two templates enumerate: resources/list returns one entry per agent and one per job, with the real domain or the real budget in the title, so a picker shows the roster rather than a URI pattern you are expected to fill in yourself. Every read hits the chain at the moment it is read — there is no cache, because a cached escrow figure is how an agent ends up doing unpaid work.

Prompts

Three guided flows, each of which is mostly about the order the tools have to be used in:

Prompt

What it walks through

vet_agent

Resolve → reputation → settlements → job history → quote, ending in a recommendation that is allowed to be "no"

post_and_fund_job

Compose post_job, then say plainly that a budget commits no money, then compose the funding group

settle_job_escrow

Read the status, pick release or refund from it, name who is about to be paid, and say who may sign and when

Every tool a prompt names exists — PROMPTS_REFERENCE_TOOLS is asserted against TOOL_NAMES in the tests. A prompt that sends a model after a tool that isn't there gets improvisation instead of a chain read, and for "has this agent been paid" that means an invented track record.


Skills

A skill is the smallest thing an agent can advertise and be paid for: a stable id, a description, an input schema, and a price. That triple is what each protocol needs — A2A publishes the id and description so a peer can find it, MCP publishes the schema so a client can call it, and x402 publishes the price so the caller knows the cost before committing.

Four ship, all backed by the registries:

Skill

Price

Backed by

ripar.identity.resolve

free

ag_ / dm_ / ad_ boxes on 769444119

ripar.reputation.report

$0.01

sc_ boxes on 769444120

ripar.settlement.audit

$0.02

sc_ boxes on 769444120 + the Algorand indexer

ripar.validation.post-job

$0.05

jb_ boxes on 769444121 (returns an unsigned tx)

Reads that cost nothing to serve are priced at zero rather than given a token price. Charging for a public box read would be theatre.

import { SKILLS, skillsManifest } from "@ripar/skills";

console.log(skillsManifest("testnet"));
// { asset: { id: 10458941, symbol: "USDC", decimals: 6 }, skills: [ … ] }

ripar.settlement.audit is the one worth looking at. It reads USDC transfers from the indexer and reports them next to the score the ReputationRegistry actually holds, so a claimed track record can be checked against money that demonstrably moved.

What it deliberately does not report is a per-transfer "already credited" flag. The registry used to keep a pd_ box per counted payment and this skill joined against it; that ledger is gone, because keying it on the txid was circular — the box name depends on the txid, which depends on the group id, which depends on the app call, which must declare the box — and unnecessary, since the payment is now a transaction in the same group and consensus rejects a duplicate. Leaving the join in place would have marked every real payment uncredited: a lie in the shape of an answer.


A2A: publishing and discovering

Publish your card

RIPAR_AGENT_ID=1 \
RIPAR_PAY_TO=UBB4PNTT7CI3IQS25ZMQR4DGVYYCBORNSBLU4WKUGX4BAZ3KN4O2KATPAU \
RIPAR_AGENT_URL=https://your-agent.example \
npm run card

Serves the card at both /.well-known/agent.json and /.well-known/agent-card.json. Or drop it into an app you already have:

import { riparAgentCard, createCardHandler } from "@ripar/skills/a2a";

const card = riparAgentCard({
  name: "My Ripar Agent",
  description: "Does a specific useful thing",
  url: "https://your-agent.example",
  agentId: 1,          // your IdentityRegistry id
  payTo: "UBB4…",      // where x402 payments land
});

const handleCard = createCardHandler(card);
// returns true if it answered, false if the path wasn't a card — so it
// sits in front of your existing routes rather than taking them over

Discover someone else's

import { discoverAgent } from "@ripar/skills/a2a";

const result = await discoverAgent("agent-1785821796525.ripar.io");

result.endpoint          // where to talk to it
result.skills            // what it says it can do
result.x402.prices       // what each skill costs
result.mcp.tools         // which MCP tools it exposes
result.verification      // whether the chain agrees with the card
result.warnings          // things worth knowing before you pay

verification is the part that matters. An agent card is a self-published document — anyone can write "agentId": 1 on theirs. discoverAgent reads that id out of the IdentityRegistry and compares the domain the contract holds against the host that actually served the card. Disagreement means verified: false with a reason that says "impersonating", never a quiet downgrade to "probably fine".


A2MCP: why these two protocols belong in one package

They're perpendicular, which is exactly why they compose.

  • MCP is agent-to-tool. A client already knows about a server and calls its tools. It answers how do I invoke this? — and has no discovery story at all. Someone has to put the server in a config file first.

  • A2A is agent-to-agent. An agent publishes a card at a well-known URL. It answers who is out there and what can they do? — and has no invocation story for tools. A card describes skills, not function signatures.

Stack them and each fills the other's hole. A Ripar card carries an MCP extension naming the exact server the agent exposes and the tools on it:

{
  "capabilities": {
    "extensions": [
      {
        "uri": "https://ripar.io/a2a/ext/mcp/v1",
        "params": {
          "transport": "stdio",
          "command": "npx",
          "args": ["-y", "@ripar/skills"],
          "tools": ["ripar_search_agents", "ripar_get_agent", "…"]
        }
      }
    ]
  }
}

So the whole path is:

GET https://them.example/.well-known/agent.json    ← A2A discovery
  → read capabilities.extensions[".../mcp/v1"]     ← which tools exist
  → connect to that MCP server                     ← MCP invocation
  → call the tool                                  ← work happens

Discovery and invocation end up one hop apart. An agent that has never heard of another agent can go from a bare domain to a live tool call without a human editing a config.

The third leg is payment. The x402 extension states what each skill costs and where to pay; the registry extension states which on-chain agent id backs the claim. So the caller can check reputation before it spends. Price, track record, and the call itself all come off one document.

Extension data rides in capabilities.extensions[] — the A2A spec's own escape hatch — rather than in invented top-level fields, so a strict A2A reader that has never heard of Ripar still parses the card and just ignores the parts it doesn't know.

A note on card paths and versions

The current A2A spec puts the card at /.well-known/agent-card.json and models transports as supportedInterfaces[]. The older, widely deployed shape used /.well-known/agent.json with flat url + preferredTransport + protocolVersion.

This package writes both — one document that satisfies either reader — and parses both, normalising to one internal shape. You don't have to know which generation a peer speaks.


Reading the chain yourself

No API key, CORS open, nothing to sign up for:

import { RiparRegistry } from "@ripar/skills";

const registry = new RiparRegistry();               // TestNet by default

await registry.totalAgents();                       // global state agent_count
await registry.listAgents();                        // every ag_ box, decoded
await registry.resolveByDomain("agent-…​.ripar.io"); // dm_ index → agent id
await registry.getScore(1);                         // sc_ box
await registry.listJobs({ status: "assigned" });    // jb_ boxes
await registry.listJobsWithEscrow({});              // jb_ boxes ⋈ es_ boxes
await registry.getEscrow(2);                        // es_ box, 0 when unfunded
await registry.escrowTerms();                       // asset, window, app account
await registry.settlements({ agentId: 1 });         // indexer, plus the score box

If the chain is unreachable these throw. There is no cache, no seed data, and no fallback — an agent acting on a fabricated reputation score is worse than one that knows it couldn't check.

About the box encoding

Box values are ARC-4 encoded structs, so they're decoded with ABIType.from(...) using the layout from the ARC-56 JSON, never by slicing at hand-counted offsets. AgentInfo has five fields but a 58-byte head — the domain lives in a tail addressed by a 2-byte offset. Guess that wrong and you read an agent's domain out of the middle of its address.

Box names are raw AVM bytes, which is a different rule:

ag_ + 8 raw big-endian bytes       uint64 agent id
dm_ + the domain's UTF-8 bytes     no ARC-4 length prefix
ad_ + the 32-byte public key       not the 58-character address string
sc_ + 8 raw big-endian bytes       uint64 agent id
jb_ + 8 raw big-endian bytes       uint64 job id
es_ + 8 raw big-endian bytes       uint64 job id — value is a bare uint64, not a struct

Both rules are pinned by tests against real captured bytes.

es_ has one more rule worth stating: an absent box is zero. The contract never creates one for a job nobody funded, and it deletes the box before it sends the money out, which is what makes paying an escrow twice impossible. So escrow 0 on a finished job means it was settled, not that it never existed.


Configuration

Everything defaults to TestNet and public endpoints. Environment variables, all optional:

Variable

Default

Notes

RIPAR_NETWORK

testnet

The registries only exist on TestNet today

RIPAR_ALGOD

https://testnet-api.algonode.cloud

RIPAR_INDEXER

https://testnet-idx.algonode.cloud

Asking for mainnet without supplying app ids fails loudly rather than reading app 0 and reporting an empty registry.


Tests

npm test              # 192 tests, including live TestNet reads
RIPAR_SKIP_LIVE=1 npm test   # offline: skips the 13 live tests
npx tsc --noEmit      # typecheck

File

Covers

test/decoders.test.ts

ARC-4 box decoding and box-name encoding, against bytes captured from the live chain

test/card.test.ts

A2A card emitting, parsing, warnings, and impersonation detection

test/tools.test.ts

Tool schemas and annotations, plus the real server over an in-memory transport

test/escrow.test.ts

Reading es_, and every refusal and resource array in the three escrow composers

test/resources.test.ts

Resources and prompts over the real MCP protocol, including that prompts name real tools

test/skills.test.ts

The four skills, their prices, and their input guards

test/x402.test.ts

402 challenge parsing and the never-invent-a-payment rule

test/payee.test.ts

The payee-check rule, the shared cache, decoding who a signed header pays, and the call gate

test/live.test.ts

Real reads against TestNet — shape invariants, not fixed values

The fixtures aren't invented. Every base64 blob in decoders.test.ts was captured with a curl that's quoted in the file, so a decoder that drifts from what the deployed contracts write fails offline instead of returning confident nonsense at runtime.

The escrow transactions were checked against the real AVM before any of them shipped, by running the composed groups through algod's simulate endpoint — which needs no key and spends no fee, and which reports a missing box or foreign app as unavailable App … and a short inner-transaction budget as group fee … too small, both of them long before the contract's own asserts are reached. Dropping the identity app from release_escrow produces exactly the first; dropping the extra fee produces exactly the second.

The tests are mutation-checked: each new assertion was verified by breaking the line it covers and confirming the suite goes red. 45 mutations, 45 killed — including grouping the funding transactions the wrong way round, paying a refund to the sender instead of the client, measuring the dispute window from the wrong timestamp, and pointing a prompt at a tool that does not exist.


Layout

src/
  config.ts        app ids, endpoints, USDC, job statuses, box prefixes
  abi.ts           ARC-4 box decoders and box-name encoders
  registry.ts      the read side — algod, indexer, escrow, and the settlement join
  unsigned.ts      the write side — composes, never signs
  x402.ts          402 challenge parsing, quote vs call
  payee.ts         the paid payee check, its 24h cache, and the call gate
  skills.ts        the four skills
  a2a/
    card.ts        build and parse agent cards (both generations)
    discover.ts    discoverAgent() + on-chain verification
    server.ts      Ripar's own card, and a handler that serves it
  mcp/
    tools.ts       the tools, defined as data
    resources.ts   ripar:// documents, including the two that enumerate
    prompts.ts     three guided flows, each naming only tools that exist
    server.ts      McpServer wiring
  bin/
    mcp-stdio.ts   the stdio entry point
    serve-card.ts  the card server

Known gaps

Things that are true about the deployed system today, written down because finding them in production is worse.

1. Escrow is denominated in a test asset, not in circulating USDC. The live ValidationRegistry is bootstrapped against ASA 768547363 (rUSDC), a six-decimal token minted for this deployment because the TestNet USDC faucet is login-gated. escrowTerms() reports that id rather than assuming it. Read the assetId a composer gives you; do not assume "USDC" because the amounts are formatted with six decimals. Moving to circulating USDC (10458941) means a redeploy — bootstrap is one-shot — and an older registry still answers with the old asset forever.

2. The volume on TestNet is a proof, not traction. Agent 1's score box currently reports one credited payment of 0.010000 and two passing validations. Those are real — the payment is an Algorand transfer that the contract read out of its own group, and the verdicts were written by the ValidationRegistry through record_validation — but they were produced by deploy-v2.mjs exercising the contracts, not by anyone buying anything. Do not quote them as usage.

3. Box listings are capped at 100 pages of 1000. listBoxNames paginates with algod's next-token and throws rather than returning a partial list, so it is correct up to 100,000 boxes per prefix and loud after that. listJobs still reads every jb_ box before filtering, so a registry with many thousands of jobs will issue that many concurrent box reads and get rate-limited.

4. npx tsc --noEmit does not typecheck test/. tsconfig.json excludes it. Four type errors in test/x402.test.ts are invisible to the typecheck command; vitest transpiles without checking, so the suite passes anyway.


License

MIT

Available Tools

8 tools
ripar_call_endpointCall an endpoint (x402-aware)A

Call an endpoint and return its response. If it answers 402, you get the payment challenge back instead of a result — this server holds no private key and CANNOT pay. To complete a paid call, have a wallet sign the payment and pass the resulting header as paymentHeader; it is forwarded untouched as X-PAYMENT.

Inputs: url (required), body, method, headers, +1 more.

ripar_get_agentGet one Ripar agentA
Read-only

Fetch a single agent record from the IdentityRegistry by id, domain, or Algorand address. Returns found:false with the reason when the registry has no such agent — the contract's 'not found' value is a literal 0, so an absent agent is a real answer, not an error.

Inputs: domain, address, agentId, includeJobs, +1 more.

ripar_get_reputationGet an agent's reputationA
Read-only

Read an agent's score from the ReputationRegistry (Algorand TestNet app 768559198): payments credited to it, total USDC volume, and validator verdicts. Each credit is keyed to a payment transaction id and the contract refuses to count the same id twice, but it does NOT verify that the id names a real transfer — so treat a score as a claim recorded on chain, not one proven by it, and call ripar_settlements to check it against the indexer before trusting a number. An agent with no score box has never been credited at all, which is different from having been paid and scored zero.

Inputs: agentId (required).

ripar_list_jobsList validated jobsA
Read-only

List jobs on the ValidationRegistry (Algorand TestNet app 768547172), newest first, optionally filtered by status or by the agent serving or validating them. Each job commits to its spec by hash; the spec and the result themselves stay offchain.

Inputs: jobId, limit, status, agentId.

ripar_post_jobCompose a post-job transaction (unsigned)A
Idempotent

Compose a ValidationRegistry post_job call and return it UNSIGNED as base64 msgpack, with a plain-language summary of what signing it would do. Nothing is submitted and no key is used or held: a human or wallet signs and broadcasts. The spec is committed by hash so it cannot be changed after the job is open.

Inputs: sender (required), specHash (required), budgetMicro (required), validatorAgentId.

ripar_quote_endpointQuote an x402 endpointA
Read-only

Ask a paid endpoint what it charges, without paying. Makes the request, reads the HTTP 402 challenge, and reports the cheapest acceptable payment: amount, asset, network, and payee. An endpoint that answers 200 is reported as free rather than as an error.

Inputs: url (required), body, method, headers.

ripar_search_agentsSearch Ripar agentsA
Read-only

List or search agents in the on-chain IdentityRegistry (Algorand TestNet app 768547159). Matches a substring of the agent's domain, or an exact agent id or Algorand address. Returns live registry records — if the chain is unreachable this fails rather than guessing.

Inputs: limit, query, withReputation.

ripar_settlementsSettlement history and reputation gapA
Read-only

List real USDC transfers for an agent from the Algorand indexer, each marked with whether the ReputationRegistry has already counted it. Inbound payments marked counted:false are reputation the agent earned but was never credited for — a gap that is only visible because the transfer log and the registry's pd_ boxes are read together.

Inputs: limit, address, agentId.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.4/5.0

    Scored across 8 tools

    Disambiguation5/5

    Each tool targets a distinct resource and action: search vs. get for agents, get for reputation, list vs. post for jobs, quote vs. call for endpoints, and settlements as a separate verification tool. There is no meaningful overlap that would confuse an agent.

    Naming Consistency5/5

    All tool names follow the same pattern: ripar_<verb>_<noun>. Verbs are consistent (search, get, list, quote, call, post) and lower_snake_case is used throughout. This creates a predictable and readable API surface.

    Tool Count5/5

    8 tools is well-scoped for the server's purpose, covering agent discovery, reputation, jobs, settlements, and endpoint interactions without unnecessary bloat or redundancy. Each tool earns its place.

    Completeness4/5

    The set covers the core workflows: reading identities and reputation, listing and creating jobs, verifying payments, and interacting with paid endpoints. Minor gaps exist (e.g., no get_job by id, no on-chain update actions), but these are acceptable given the server's read-only and transaction-construction role.

    Related MCP Connectors

    Related MCP Servers