Underdog by Conway Research

# Fix list: Underdog

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/underdog, the October 2026 research run, assessed 3 October 2026. Grade F, 29.9 out of 100.

This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public.

For a coding agent working on Underdog: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published.

## 1. Security & auth, 14 out of 100, up to 15.1 more on the total

Why it scored 14: Read as local software that holds the owner's mail and calendar, with the five framework lines the calibration CLIs used (what leaves the machine and credentials, approvals, guardrails, audit, security programme). Conway says Woof runs on the owner's machine "with nothing sent anywhere" and that Underdog runs "even without wifi" (husky.underdog.ai), and the models need no account. How mail and calendar access is granted, stored and revoked, and what the app itself sends, would be in a privacy policy on underdog.ai, which refuses our reader, so that part scores as absent because we couldn't read it (10 of 30). Woof 4B and 2B 1.1 are described as DOM browser executors in their release files, and the husky-flash card lists tool calls, mail and browser steps among Underdog's prompts. We found nothing on approval before it sends or acts, and underdog.ai is unchecked (0 of 20). It reads mail and web pages, which are untrusted content, and we found no prompt-injection guidance. Conway's PII redaction model is a privacy filter and isn't said to ship in Underdog (0 of 15). No audit or per-action log described (0 of 15). No security.txt at conway.tech (404), no SECURITY.md in Conway's three GitHub repositories, and no disclosure policy, advisories or bug bounty found. underdog.ai's security.txt is unchecked. Weights are safetensors, and Woof 4B and 2B 1.1 publish a SHA-256 for each file (4 of 20). `husky serve` has no documented interface, so it adds nothing here.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-security):

- 0 to 30, the credential model. 30 for OAuth 2.1 with scopes, or scoped and revocable keys with rotation. 20 for plain revocable API keys. 10 for one all-powerful key. 10 off when a secret can travel in a URL query string as a documented option.
- 0 to 20, read-only or least-privilege modes, and confirmation or approval for destructive actions.
- 0 to 15, prompt-injection posture where the tool returns untrusted content (documented mitigations or guidance). A tool that returns no untrusted content gets 10.
- 0 to 15, audit logs or per-call visibility for the operator.
- 0 to 20, a security programme. security.txt or a disclosure policy, a bug bounty, SOC 2 or ISO 27001, advisories handled in public.

Models are read for retention, whether API data trains models (and whether that's off by default), zero-retention options and certifications. Frameworks for telemetry defaults, approval hooks, guardrails and sandboxing.

## 2. Agent ergonomics, 15 out of 100, up to 13.8 more on the total

Why it scored 15: Read for the models, since we found no API, MCP server, CLI or SDK for the app, and underdog.ai, where one would be documented, refuses our reader. The husky-flash card's `husky serve --model ConwayResearch/husky-flash` comes from an "Underdog Greyhound repository" that isn't public, and no card or page gives its port, protocol or install, so it earns nothing here (checked 3 October at 08:20 UTC). No card states a context length. The husky-flash card times a continued conversation at 3.6 to 5.5 times MLX's speed to the first word, which we read as prompt reuse. Judgement call (5 of 25). The 27B cards' `--default-reasoning-effort` and the ternary card's `--language-only` are the only output controls we found (3 of 20). No documented errors (0). Nothing on retries or statelessness (0). The 27B cards recommend medium reasoning and say Splash prepares the draft model itself, Woof 1.0 loads in Apple's mlx-lm and mlx-swift, and the ternary card documents an OpenAI-compatible API through Splash, a third party's server, so standard OpenAI clients reach the 27B model. No Conway SDK (7 of 15).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-ergonomics):

- 0 to 25, context cost. For MCP, the number and size of the tool definitions (25 for ten or fewer compact tools, 15 for 11 to 30, 5 for more than 30, plus up to 10 back for toolsets, dynamic loading or read-only subsets). For APIs, whether responses can be sized (field selection, limits, summaries).
- 20, pagination, filtering and output-size controls.
- 20, actionable, documented error responses, codes and messages an agent can recover from.
- 20, idempotency or safe retries, and for MCP the `readOnlyHint` and `destructiveHint` annotations.
- 15, sensible defaults, few required parameters, and official SDKs in at least two languages.

Models are read for tool use, structured output, prompt caching, context length, batch and SDKs. Frameworks for how much code and how many defaults a tool-calling agent with MCP needs.

## 3. Reliability, 33 out of 100, up to 13.4 more on the total

Why it scored 33: Read with the local-software lines, since Underdog is an app the owner runs with no hosted service behind it for an agent, the reading the calibration CLIs and model used. The app's download page, macOS requirements and signing are on underdog.ai, whose robots.txt refuses our reader, so the app earns nothing on this line because we couldn't read its site, not because something is missing. The models install from Conway's own Hugging Face organisation as MLX builds, with runtimes named on two cards (MLX 0.32.0, MLX-LM 0.31.3 and Splash 1.1.0 or later on the 27B card, mlx-lm and mlx-swift on the Woof 1.0 card) and a SHA-256 for every file of Woof 4B and 2B 1.1 (10 of 20). No public CI or test suite, since the app's source isn't published and none of Conway's three public GitHub repositories mentions Underdog, Woof, Bark or Husky (0). No public tracker. The discussion tabs on Woof 4B 1.1 (1,185 downloads in the last month) and Underdog 27B 1.0 had no threads on 3 October, so there's nothing to count, and the app's support pages are unchecked. Judgement call (5 of 25). The models carry major.minor versions (Woof 1.0 on 4 September, 1.1 on 12 September) with no changelog saying what changed, and the app's release notes are unchecked (3 of 15). Woof 4B and 2B 1.1, Bark 0.8B 1.0 and Underdog 27B 1.0 are all at 1.0 or later. We read this line for the models, as the install and changelog lines were, and the app's own version is unchecked (15).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-reliability):

Hosted APIs, MCP servers, models and platforms.

- 20, a public status page with component history (Statuspage, Instatus, BetterStack or the vendor's own).
- 0 to 30, the incident record for the last 90 days on that page. 30 for a clean record or trivial incidents only, 20 for minor incidents only, 10 for one major outage (an hour or more of a core API down, or errors across the board), 0 for several. 5 when there's no history we could read, and the note says so.
- 15, rate limits documented with numbers.
- 15, documented 429 or overload handling (Retry-After, backoff guidance), and idempotency keys or safe-retry guidance where writes are involved.
- 10, an SLA published for any paid tier.
- 10, the surface agents use is generally available, not beta or preview.

Local packages, SDKs, frameworks and stdio MCP servers.

- 20, installs from an official package with supported runtimes stated.
- 25, a public CI and test suite, passing on the default branch.
- 0 to 25, open crash or regression issues relative to activity (25 for few and handled, 0 for many, old and unanswered).
- 15, semver discipline and breaking changes called out in a changelog.
- 15, version 1.0 or later, or declared stable.

Protocols are read from their reference implementations, the public facilitators or servers, spec stability and test vectors.

## 4. Schema & documentation, 24 out of 100, up to 12.4 more on the total

Why it scored 24: Read for a platform whose models are part of the product, so the model cards are the documentation, as the calibration model's card was. No API reference or OpenAPI file from Conway. The repositories ship config.json and a chat template, and husky-flash a manifest.json, that runtimes read, and the 27B ternary card says Splash (Inco AI's engine, not Conway's) takes requests through an OpenAI-compatible API, with no spec from Conway (5 of 25). No llms.txt at conway.tech (404 on 3 October), and underdog.ai's is unchecked. The model cards are Markdown on Hugging Face (5 of 10). The 27B cards state the model's purpose (conversation, writing, coding and everyday assistance), the woof-2B-mlx-4bit-v1.1 card says it's for browser tool use and names its inputs and outputs, and the husky-flash card explains each part of its package. The Woof 4B 1.1, Woof 2B 1.1 and Bark 0.8B 1.0 cards are one or two sentences, and no card says when not to use a model (6 of 20). Nothing Conway publishes for a caller has typed inputs, and no card states a context length or input limit (0 of 15). Run commands on the 27B, 27B ternary and husky-flash cards, none on the Woof 1.1 or Bark cards, and no documented errors (4 of 15). Versioned model names and public Hugging Face commit history, with no changelog or release notes we could read (4 of 15). The app's own documentation would be on underdog.ai, unchecked.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-schema):

APIs and MCP servers.

- 25, a machine-readable contract (a public OpenAPI file or similar; for MCP, typed JSON Schema inputs on every tool).
- 10, llms.txt or Markdown docs served for agents.
- 0 to 20, descriptions that say what a tool is for, when to use it and when not to, read from the tool definitions in the source or the API reference.
- 0 to 15, typed inputs with enums, constraints and required fields, and no free-form JSON blobs.
- 0 to 15, examples and documented error responses.
- 15, versioning and a public changelog.

Models are read from the API reference, the OpenAPI file, llms.txt, the structured-output and tool-use docs and the model cards. Frameworks from docs a model can follow, typed interfaces, examples and the API reference.

## 5. Transparency & trust, 24 out of 100, up to 6.7 more on the total

Made of editorial 23, provenance 24.

Why it scored 24: The editorial half. Underdog 27B 1.0 and its ternary build, Woof 4B and 2B 1.1 and Bark 0.8B 1.0 are Apache-2.0 with a `LICENSE` file, woof-1.0-4B carries only the tag, and husky-flash is marked other with a card saying Woof's licence applies. The Woof 1.1 and Bark cards don't name their base models (Woof 2B 1.1's `LICENSE` names OpenBMB, and the 27B cards say Qwen3.8). The weights' licences are clear. The app's source isn't published and its terms are on underdog.ai, unchecked, so the app adds nothing (15 of 30). Conway's statements agree with each other that the models run on the owner's machine, send nothing and work offline (husky.underdog.ai, the founder's page), but none says what the app stores, for how long or how to delete it, and the privacy policy is on underdog.ai, unchecked (8 of 30). No deprecation policy or dated notices found, and woof-1.0-4B stays up beside 1.1 with nothing saying which to use (0 of 20). We found no statement on app telemetry, crash reports or update checks, and no opt-out. "Nothing sent anywhere" describes Woof's inference and is credited above, and the settings and policy that would cover telemetry are on underdog.ai, unchecked (0 of 20). The provenance half is computed from the provenance block, where terms, privacy and security.txt are unchecked.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-transparency):

- 0 to 30, source availability and licence clarity. 30 for open source under an OSI licence, 15 for closed with clear terms, 0 for unclear terms.
- 0 to 30, data handling and retention statements that agree with each other (privacy policy, DPA, retention periods, subprocessors).
- 0 to 20, a deprecation policy or notices with dates.
- 0 to 20, telemetry disclosed with an opt-out (local software), or subprocessors and data locations disclosed (hosted).

The other half of Transparency and trust is the provenance score, computed from checked facts (below). The category score is the mean of the two.

Provenance checks not met in full (half of this category, computed from checked facts):

- Legal entity named: not found (0 of 20)
- Domain age: underdog.ai, no registry record we could read (0 of 15)
- Status page: not found (0 of 10)
- Changelog: not found (0 of 10)
- security.txt: could not be fetched (0 of 10)

## 6. Payments & pricing, 60 out of 100, up to 5 more on the total

Why it scored 60: Self-hosted rule, scored for the app. No payment protocol (0). Underdog's pricing page says "100% free", "Free forever? Yes. It's your computer doing the work" and "There is nothing to bill you for", so the app is free with nothing to buy, which earns 20 + 20 + 20 under the self-hosted rule, as it does for every free package run on the owner's machine. Our reader is refused by underdog.ai's robots.txt, so the page's text was supplied to us on 3 October 2026 and counted with that as its source. The page also links a sign-in, and whether the app needs an account is unchecked.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-payments):

The published rubric, also on the [x402 page](https://www.anchorterminal.com/x402/).

- 40, a machine payment protocol (x402, MPP or L402) on the tool's own endpoints. 10 to 30 when it covers only some endpoints or only goes through a third party, and the note says which.
- 20, per-call or per-unit pricing published without a login. 10 for public plan-only pricing, 0 for "contact sales" or prices behind a login.
- 20, a free tier or trial that doesn't need a card.
- 20, autonomous onboarding, meaning an agent can get access without a person signing up in a browser (keyless use, x402, a programmatic key API).

Payment platforms and agent wallets rarely charge for their own API over a machine protocol, so the first line has steps for them, and the highest one that applies counts. 40 when x402, MPP or L402 runs on all their own endpoints, 30 when it runs on part of their own API, 25 when their merchants can accept one, 20 for running a facilitator, 15 for paying as a buyer, and 0 when the only protocol is their own. Merchant acceptance sits above a facilitator because the platform's own customers can charge agents through it, while a facilitator settles for sellers who wire up the protocol themselves. The counter-argument (a facilitator does more for the protocol as a whole) has a point. Each note says which step applied.

Open-source software you run yourself is scored on its hosted or paid option if it has one. A free, self-hosted package with nothing to buy gets 20, 20 and 20 for the last three lines, and 0 to 40 for the first only if it ships a payment protocol.

## 7. Maintenance & community, 56 out of 100, up to 3.9 more on the total

Why it scored 56: Models counted as releases, as for the calibration model. Underdog 27B 1.0 and its ternary build went up on Hugging Face on 30 September 2026 and were last updated on 1 October (30). Eight Underdog model repositories between 4 and 30 September (woof-1.0-4B on 4 September, Woof 4B 1.1, Woof 2B 1.1, Bark 0.8B 1.0 and woof-2B-mlx-4bit-v1.1 on 12 September, husky-flash on 21 September, Underdog 27B 1.0 and its ternary build on 30 September) (20). Closed-source app, so the 0 to 15 scale for closed services applies. No public changelog, a contact address on conway.tech (root@conway.tech) that we didn't test, and no threads on the discussion tabs of Woof 4B 1.1 or Underdog 27B 1.0, so we saw no channel answer. Support pages on underdog.ai are unchecked (2 of 15). No official SDK and no MCP registry entry (0). The 27B card is tested against current MLX 0.32.0 and MLX-LM 0.31.3, and Woof 4B and 2B 1.1 list file hashes, with no visible CI (4 of 10).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-maintenance):

- 0 to 30, time since the last release, or the last published model or API change for a closed service. 30 within 30 days, 20 within 90, 10 within 180, 0 older.
- 20, at least three releases or dated changelog entries in the last 90 days.
- 0 to 25, responsiveness. Issues and pull requests answered on GitHub (the open issues and how recent the replies are). For closed services, a public changelog and a support or community channel that answers, 0 to 15.
- 15, presence in the official MCP registry under a verified namespace (MCP servers), or current official SDKs (APIs and models).
- 10, package health, current dependencies and CI.

Models are read for deprecation notice periods and model churn rather than release counts.

## What we couldn't check

What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it.

- unchecked: underdog.ai terms of service (its robots.txt refused our reader on 3 October 2026)
- unchecked: underdog.ai privacy policy, retention, what the app stores about mail and calendar, and what it sends off the Mac
- unchecked: telemetry, crash reports and update checks in the app, and whether they can be turned off
- unchecked: the download page (https://underdog.ai/download), installer signing, system requirements, and whether the app needs an account (the pricing page links a sign-in)
- unchecked: the app's release notes or changelog, its version, and whether it's declared stable
- unchecked: a security contact or security.txt on underdog.ai (conway.tech/.well-known/security.txt is a 404)
- unchecked: any llms.txt, API, MCP server or CLI documentation on underdog.ai
- unchecked: underdog.ai's domain registration date and the legal entity. a16z and GitHub say Conway Research, conway.tech says Conway, the automaton licence reads 'Copyright (c) 2026 Conway', and we found no registered company name
- What `husky serve` exposes. The husky-flash card says the Husky server comes from the 'Underdog Greyhound repository', which isn't under Conway-Research on GitHub (no greyhound, husky or underdog repository there on 3 October) and which we found nowhere else, and no page gives its port, protocol or install. Scored as no interface until Conway documents one
- Investors. a16z's own announcement of 2 October 2026 says it led Conway Research's first round and names no other investor. It mentions Anthropic only in the founder's background (an Anthropic co-founder invited him to test an early Claude). We found no first-hand source that Anthropic invested, so the news report that it did is unchecked. Our graders run on Anthropic's Claude models, so an Anthropic stake would belong in the listing's disclosures
- Conway's home page says Underdog 27B beats Claude Opus 4.6. The 27B cards publish speed figures only, and any quality evaluation would be on underdog.ai, unchecked. Judgement call, since our graders run on Claude models. We neither scored nor deducted for the claim
- Whether Underdog keeps or searches a memory of the owner. Conway's pages say it connects mail and calendar and nothing we could read says more, so memory.user and memory.search were taken out of capabilities, and they should come back if underdog.ai confirms them
- Whether Underdog acts in a browser. Woof 4B and 2B 1.1 are described as DOM browser executors and the husky-flash card lists browser steps among Underdog's prompts, but no page we could read says what the app does in a browser, so browser.control isn't in capabilities
- Whether Underdog runs on iPhone. The Husky page says it runs on the Mac and iPhone and conway.tech says Underdog 27B runs on a Mac or iPhone, while the founder's page and the Husky page's launch note name only the Mac. The download page is unchecked
- The base models of Woof 4B 1.1 (architecture qwen3_5) and Woof 2B 1.1 (architecture llama, `LICENSE` naming OpenBMB) aren't named on their cards, so the licence chain behind the Apache-2.0 weights isn't stated
- Whether Conway's PII redaction model (pii-30M-mlx-4bit-v1.0, for browser observations) or dsv4-flash-0731-vision (no licence tag) ships in Underdog
- Hugging Face's organisation page showed fewer models to our readers than its API, which listed ten on 3 October. We used the API
- husky.underdog.ai is a separate host from underdog.ai, linked from conway.tech and the founder's page, and it loaded for our reader. We used it as a Conway page and used no caches, mirrors or proxies

## Weaknesses

- No API, MCP server, CLI or SDK of Conway's for agents, and the `husky serve` command on the husky-flash card comes from a repository that isn't public
- The app's source isn't published, and none of Conway's three public GitHub repositories mentions Underdog
- The Woof 4B 1.1, Woof 2B 1.1 and Bark 0.8B 1.0 cards are one or two sentences, with no base model, context length or limits
- No security.txt at conway.tech (404), no SECURITY.md, and no disclosure policy, advisories or bug bounty found
- Conway's home page says Underdog 27B beats Claude Opus 4.6, and the 27B cards publish speed figures only

## What costs an agent a turn today

The notes we give agents before they call it. Each one is a workaround an agent shouldn't need.

- Don't look for an agent interface to Underdog. We found no API, MCP server, CLI or SDK, and underdog.ai, where one would be documented, refuses our reader
- Don't count on `husky serve`. The husky-flash card names it, but the Greyhound repository it comes from isn't public and no port or protocol is documented
- Run `splash serve --model ConwayResearch/Underdog-27B-1.0 --default-reasoning-effort medium` with Inco AI's Splash 1.1.0 or later for an OpenAI-compatible endpoint on `127.0.0.1:8000`, and pass `--api-key`, since Splash starts without authentication
- Pin a Hugging Face revision. Woof 4B went from 1.0 to 1.1 in eight days with no note of what changed
- Check Woof 4B and 2B 1.1 files against the SHA-256 values in release-provenance.json before loading them

## What the review panel asked for

- State context length on cards
- Publish husky serve source and port
- confirmation before outbound actions
- per-action audit log

## When it's done

Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.