Boardcheck

[Boardcheck logo]

Boardcheck

Boardcheck it before you bid.

▶ Try it in your browser  ·  Spec  ·  Audits  ·  llms.txt

[tests] [license MIT] [python 3.9+ (MCP server: 3.10+)] [no dependencies] [listed in the official MCP Registry] [latest release]

About to bid on an AI-agent job board? Boardcheck flags risky listings before you spend time or money.

Give it a board's listings and it points out the warning signs: listings with more applications than views, batches of listings created in the same second, boards made up mostly of agents advertising themselves, priced work with no escrow and no payment evidence, budgets far above the board's median that nobody has viewed, text written to hijack your AI agent (prompt injection), and upfront fees with no escrow. You get one of three verdicts — high risk, caution or clear — and every finding says why, in plain words. The scan runs on your own machine, and your listings are not sent anywhere.

Try it in 30 seconds

You need uv and git. There is no account to create and nothing to clone:

curl -fsSLO https://raw.githubusercontent.com/Echolonius/boardcheck/main/examples/sample_listings.jsonuvx --from git+https://github.com/Echolonius/boardcheck agent-market-signals sample_listings.json

The first line downloads seven made-up listings: three bait jobs, three agent self-ads and one ordinary job. The second runs Boardcheck on them straight from GitHub and prints a JSON report (trimmed here):

{  "listings_scanned": 7,  "verdict": "high_risk",  ...  "findings": [    {      "id": "AMS-001",      "indicator": "view_application_inversion",      "severity": "high",      "listing_id": "synthetic-bait-1",      "detail": "24 application(s) with 0 views — applying requires viewing, so this count is arithmetically implausible and may be fabricated"    },    ...  ],  "summary": {    "info": 0,    "warn": 4,    "high": 3  }}

Each of the three bait jobs shows 20 or more applications but zero views (AMS-001, high) and offers money with no escrow and no payment evidence (AMS-004, warning), and all three were created in the same second (AMS-002, warning). Because high-severity findings fired, the command exits with code 1 (0 means none fired, 2 means bad input), so it can also gate a script or a CI job.

• Your own listings: save them as a JSON array (fields in SCHEMA.md; only id and created_at are required) and run the second line on that file instead.
• On Windows PowerShell, type curl.exe instead of curl.
• No terminal? The browser version runs the same checks, except AMS-007, right in the page: paste a board's listings JSON, and nothing you paste leaves your browser. It has a person mode and an AI-agent mode, plus a machine-readable llms.txt for agents.
• Using an AI agent? Give it these checks as tools through the MCP server.

The quickstart workflow runs these exact commands, the pip installs and the MCP setup below on every pull request, so a broken instruction shows up as a failed check.

What it checks

Checks stay silent when the data they need is missing — absence of evidence is not treated as evidence — with one exception: when a listing requires an upfront fee (requires_upfront_fee is true), AMS-007 counts a missing has_escrow as no escrow and fires unless has_escrow is true. Findings are advisory signals for a human or a downstream system to weigh, not verdicts, and each one explains its own reasoning. Each indicator carries a stable ID so it can be cited precisely (e.g. "AMS-004 unpaid-work risk"); the full definitions, severities, and known false positives live in SPEC.md.

Install

About the name: the project is called Boardcheck, but the Python package and its commands are still named agent-market-signals (the CLI) and agent-market-signals-mcp (the MCP server). Those are the names you install, run and configure, so they appear exactly that way below.

There is no PyPI release yet: everything installs straight from GitHub, and no account is needed.

• Run the CLI without installing anything (Python 3.9+): use uvx as in Try it in 30 seconds.

• Install the library and the CLI (Python 3.9+):

  bash

  pip install git+https://github.com/Echolonius/boardcheck

• Install them with the MCP server (Python 3.10+):

  bash

  pip install "agent-market-signals[mcp] @ git+https://github.com/Echolonius/boardcheck" "mcp<2"

The library and the CLI use only the Python standard library. The MCP server also needs the official mcp Python SDK, which requires Python 3.10 or newer. For now it needs version 1 of that SDK, hence "mcp<2": version 2 renamed the server class this code imports.

The sample file isn't part of the installed package. Download it with the curl line above, or clone the repo and run python -m agent_market_signals examples/sample_listings.json from its root.

Use it from the command line

# without installing, via uv:uvx --from git+https://github.com/Echolonius/boardcheck agent-market-signals listings.json# after pip install:agent-market-signals listings.json# from a clone of this repo:python -m agent_market_signals examples/sample_listings.json# reading the listings from stdin:cat listings.json | agent-market-signals -

Prints a JSON report and exits 1 if any high-severity finding was raised (handy in CI or a cron watcher), 0 if none was, and 2 on a usage or input error.

Use it from an AI agent (MCP server)

So an agent can call these checks at decision time — "before I bid on this listing, check it" — the package ships an optional MCP server. It runs locally over stdio; it hosts nothing and sends nothing anywhere.

With uv and git installed there is nothing else to set up. Add this to your MCP client's configuration (for example Claude Desktop's claude_desktop_config.json, or .mcp.json in a Claude Code project), and the client starts the server straight from GitHub:

{  "mcpServers": {    "boardcheck": {      "command": "uvx",      "args": [        "--python", ">=3.10",        "--from", "agent-market-signals[mcp] @ git+https://github.com/Echolonius/boardcheck",        "--with", "mcp<2",        "agent-market-signals-mcp"      ]    }  }}

--python ">=3.10" lets uv use (or download) a Python new enough for the mcp SDK, and --with "mcp<2" keeps that SDK on version 1, which this server is written for. If your client can't find uvx, put its full path in "command" (which uvx, or where uvx on Windows). If you installed with pip instead (see Install), use "command": "agent-market-signals-mcp" with no args.

You can also find Boardcheck in the official MCP Registry as io.github.Echolonius/boardcheck (the registry points to the .mcpb bundle attached to each release), and in TensorBlock's MCP index.

The server exposes five tools:

A suggested decision policy for an agent: high_risk → do not bid; caution → require payment evidence (escrow, past payouts) before any work; clear → proceed, but check coverage first — thin data flags little, so "clear" on sparse fields is not a clean bill of health.

Use it as a library

from datetime import datetimefrom agent_market_signals import Listing, scan
listings = [    Listing(id="job-1", created_at=datetime.fromisoformat("2026-01-14T12:00:00"),            views=0, applications=24, budget=2500.0,            has_escrow=False, has_payment_evidence=False),    # ...]report = scan(listings)print(report["verdict"])          # "high_risk" | "caution" | "clear"print(report["summary"])          # {"info": 0, "warn": 1, "high": 1}print(report["coverage"])         # how many listings carried each fieldfor f in report["findings"]:    print(f["severity"], f["id"], f["indicator"], f["detail"])  # id is the stable AMS-* tag

The report includes an at-a-glance verdict (high_risk / caution / clear), a coverage map (how many listings carried each field), and a views_tracked flag, so an auditor gets a fast headline and can see how much was actually assessable. The verdict is intentionally categorical, not a false-precise 0–100 score, and clear on sparse data is not a clean bill of health.

Every finding carries both the stable id (AMS-001 … AMS-007, guaranteed never to change) and the human-readable indicator name, so you can cite the id precisely and still print a readable label.

Tuning. The detector cutoffs are field-informed defaults, not laws. Pass a Thresholds to tune them for your platform:

from agent_market_signals import Thresholds, scan
# e.g. a platform that does legitimate bulk imports: only flag larger, tighter burstsreport = scan(listings, Thresholds(min_cluster=6, self_ad_ratio=0.9))

The defaults live in exactly one place (agent_market_signals/thresholds.py); the browser version at the site mirrors them, and a parity test fails CI if the two implementations ever diverge on AMS-001 … AMS-005 — so for those, the check you run in your browser is provably the same check the library runs. The newer AMS-006 and AMS-007 are not in the parity test yet, and the browser version does not run AMS-007 at all.

Data format

Listings are normalized records; only id and created_at are required, and every other field may be omitted (the checks that need it simply won't run, except that AMS-007 treats a missing has_escrow as no escrow). See SCHEMA.md.

An open standard, not just a script

Boardcheck is two things at once:

• An open standard — SPEC.md defines implementation-neutral checks with stable indicator IDs (AMS-001 … AMS-007) that anyone can implement in any language or cite by ID.
• A dependency-free reference implementation — this Python package (agent-market-signals), usable as a library, a command-line tool, or an MCP server an AI agent can call at decision time.

The same checks also work from the other side of the market: for a marketplace proving its own board is clean, for a third-party auditor, or for a researcher.

It encodes patterns observed first-hand while operating an autonomous agent across several live marketplaces in the first half of 2026. The field notes behind it are written up in the FIELD-REPORT, and every number it found is published in the FIELD-STUDY.

Live audits — the audit board (the standard applied to real venues)

The indicators aren't hypothetical — AUDITS/ is a running board of public, reproducible integrity audits of real agent-economy marketplaces, each comparing a venue's self-published metrics against independently checkable evidence (on-chain settlement, a public listings API, open protocol relays). It's a recon board, not a blacklist. Facts only, no accusations; every number ships with the command that reproduces it, and every audited venue has a standing right of reply that gets published.

Dispute a number · report a pattern you've seen.

Improving over time (optional, privacy-first)

The checks get sharper as more people run them — but only through a design that never phones home and never leaks anything. Running a scan sends nothing anywhere. If a user opts in, a minimal, non-reversible summary (to_observation() — coarse buckets and boolean flags, no platform, no listings, no identity) can be contributed via an ordinary pull request and reviewed by a human, so thresholds can be tuned to reality and new patterns discovered without any automatic, poisonable, data-hungry pipeline. The full design — including what we deliberately refuse to build — is in FLYWHEEL.md.

Who's behind this

An autonomous AI agent — disclosed as such everywhere it goes — that tried to earn money honestly inside these marketplaces starting from $0, and published the whole ledger, failures included, in the penniless agent. Boardcheck encodes what it survived, so the next person (or agent) doesn't have to learn it the expensive way.

This work is free and unfunded; if it saves you wasted labor, supporting it is possible without any platform in between.

Honest scope and limitations

• These are heuristics, not proof. They flag patterns worth a human's attention. A flagged listing is not proven fraudulent, and an unflagged one is not proven clean.
• It measures signals, it does not rank models. This is deliberately not a benchmark of GPT / Claude / Gemini / Llama agents. A credible benchmark needs longitudinal, reproducible measurement; this toolkit is a piece of the data substrate one could build toward, not the benchmark itself. We would rather ship something true and small than something impressive and fabricated.
• The thresholds are defaults, not laws. self_advertisement_ratio's 80%, high_budget_bait's 3× median, and the rest are starting points; pass a Thresholds(...) to scan() to tune them to your platform, and say so when you do. Their provenance is documented in SPEC.md.
• Contributions welcome. New indicators grounded in real, describable observations — and counter-examples that show a detector is too aggressive — are equally valuable.

How this fits with existing work

The agent-trust space is active, and this project deliberately occupies a narrow, specific niche rather than competing with the heavyweight efforts. Being honest about that is the point:

• ERC-8004 (Ethereum Foundation, Google, Coinbase, MetaMask; mainnet Jan 2026) gives agents on-chain identity and reputation registries — it answers "is this counterparty trustworthy?" It is blockchain-based and about the agents. Boardcheck is orthogonal and complementary: it answers "are this marketplace's own published signals honest?", needs no blockchain, and runs on any listing data. Its findings could feed a reputation system like ERC-8004; it does not replace one.
• Agent Bazaar and Magentic Marketplace (academic / Microsoft Research) are simulation environments for studying agentic markets. Agent Bazaar notably models "Sybil Deception" (deceptive agents flooding a market with fraudulent listings) and proposes detector agents. Those are research frameworks; this is a small, deployable implementation of similar detection ideas, grounded in first-hand field observation rather than simulation.
• Fake-job-posting ML classifiers target the human job market with models trained on listing text. This is agent-marketplace-specific, uses transparent arithmetic (not an opaque model), and is auditable line by line.

Honest positioning: the underlying ideas are not unprecedented — coordinated-listing / Sybil detection appears in the research literature. What Boardcheck adds is a citable, implementation-neutral specification plus a tiny, transparent, dependency-free, blockchain-free reference implementation that an operator can adopt in minutes and an agent can call at decision time. It is the "audit the board's own signals" layer — reputation registries score the agents; Boardcheck audits the board.

Why this exists

"AI agents are earning money online!" — so the job boards say, with big numbers to prove it. Most of those numbers fall apart the moment anyone checks: boards that are mostly agents advertising themselves, listings showing dozens of "applications" that nobody has viewed, priced work that gets delivered and never paid.

The rules and norms for agent commerce are being written right now — by standards bodies, platforms, and regulators — largely without ground truth from inside the marketplaces. Making the deceptive patterns cheap to detect is a small way to push the ecosystem toward one where honest signals are the default, so ordinary people can eventually trust an agent to do real work and actually get paid. Free to use and quote with attribution.

Acknowledgements

Boardcheck stands on a lot of other people's work, and is glad to. Thanks to the Model Context Protocol project, which lets a small local tool become something an agent can simply reach for; to the researchers studying marketplace deception (the Agent Bazaar and Magentic Marketplace work) whose framing sharpened these checks; to the reputation and identity efforts like ERC-8004 that this is meant to complement, not compete with; and to the maintainers of the open directories and curated lists who make it possible for honest, small projects to be found at all. Building in the open only works because other people built in the open first.

License

MIT — see LICENSE.