Head to head · Local inference · October 2026 research run

screenpipe vs Underdog

screenpipe scores 60.5 (C) on agent readiness against Underdog's 18.2 (F), and leads in 6 of 7 scored categories. Underdog leads on payments & pricing. Both do local inference.

Best local AI models and assistants · All 284 local ai comparisons

Which one, for what

screenpipe C

Best for An owner who wants an agent to search their own screen text, audio transcripts and meetings on one machine through MCP.

Ahead on

  • Reliability, 53 against 0
  • Schema & documentation, 83 against 0
  • Agent ergonomics, 75 against 0
  • Security & auth, 55 against 3
  • Maintenance & community, 69 against 56
  • Transparency & trust, 69 against 17

Also in its favour

  • Runs on your own machine

Watch for

The 34 tool definitions run to about 36,000 characters in the source, with no toolsets, and only the HTTP transport's three tools as a smaller set.

Underdog F

Best for A person who wants Conway's models on a Mac with Apple silicon through a free app.

Ahead on

  • Payments & pricing, 90 against 60

Also in its favour

  • No key needed to call it
  • No incidents deducted, where screenpipe loses 5 points for them

Watch for

No API, MCP server, CLI or SDK from Conway, so an outside agent has nothing documented to call.

Why the edge

Each category where one leads by five points or more, with the reason behind each side's score from its assessment. The numbers are in the table below; this is what they stand for.

Schema & documentation screenpipe +83

screenpipe 83

Read with the MCP and API lines. Every one of the 34 stdio MCP tools has a typed JSON Schema input, and the local REST API has an OpenAPI 3.0.3 file in the docs source with 59 paths, 67 operations and typed request bodies, which docs.json uses for the API reference. The file declares no security scheme and covers only part of the server's routes (25). llms.txt at docs.screenpipe.com and Markdown copies of the docs pages, per the 3 October check, not re-read on 11 November because Screenpipe's terms bar robots on the docs host (10). Tool descriptions say what each tool is for, when to use it and when not to (search-content names activity-summary and search-elements as the alternatives) and how to size a call (18 of 20). Inputs use enums (content_type has eight values), defaults and required fields, and the workflow tools set minimum, maximum and additionalProperties, but several limits sit only in prose (limit says max 20), speaker IDs and tags are comma-separated strings, and send-notification takes a free-form context object (11 of 15). curl examples throughout the docs and 401 and 403 guidance on the getting-started page, but the OpenAPI file documents only 200 responses for all 67 operations, and the MCP docs page still describes two tools (8 of 15). Version tags per component and a dated changelog in the docs whose latest section, October 2026, covers app releases to 9 October and points to the repository after that, with the OpenAPI version still 1.0.0 and no versioned path. Partly met, so the earlier 11 stays (11 of 15). 25 + 10 + 18 + 11 + 8 + 11 = 83.

Underdog 0

No interface an outside agent can call was found from Conway (no API, MCP server, CLI or SDK), so the interface lines score 0, as for fresha-ai-concierge and lola. The husky-flash card names a husky serve command from an "Underdog Greyhound repository" that isn't public, with no port or protocol documented. The Underdog app downloads an unmodified llama.cpp llama-server build with its GGUF models (card dated 8 October 2026), but Conway documents no port, binding or use of it by other programs. No llms.txt at conway.tech (404 on 11 November) and underdog.ai unread (0). Model cards with run commands describe the weights, not the app (0). 0 + 0 + 0 + 0 + 0 + 0 = 0.

Agent ergonomics screenpipe +75

screenpipe 75

Read with the MCP lines for screenpipe-mcp 0.20.4, the surface the docs connect agents to. 34 tools over stdio, 38 with a team token, in about 36,000 characters of definitions in the source, with no toolsets or dynamic loading (5). The same package's HTTP transport (npx -y screenpipe-mcp --http) registers three read-only tools, search_content, search-chats and screenpipe-skills, read as a read-only subset (5 back) (10 of 25). Searches take limit and offset, time ranges, app, window, speaker and tag filters and max_content_length, per-result text is cut to 1,000 characters by default, and screenshots stay out unless asked for (20). Tool errors return isError with the HTTP status and the next step, such as "do not treat this as an empty catalog", and the engine answers out-of-plan history reads with history_access_limited, but there's no error reference. Partly met, so the earlier 16 stays (16 of 20). All 34 tools carry annotations, 22 with readOnlyHint, merge-speakers with destructiveHint and idempotentHint on the reads, with no idempotency keys on writes such as create-pipe. Partly met, so the earlier 18 stays (18 of 20). search-content needs no parameters and its defaults suit a first call, and @screenpipe/sdk is the one official SDK, in JavaScript. Partly met, so the earlier 11 stays (11 of 15). 10 + 20 + 16 + 18 + 11 = 75.

Underdog 0

No outside-agent interface from Conway, so every line scores 0 (fresha-ai-concierge, lola). Running Conway's weights through llama.cpp, Splash or Ollama is those engines' interface, not Underdog's. 0 + 0 + 0 + 0 + 0 = 0.

Reliability screenpipe +53

screenpipe 53

Read with the local-software lines, since the desktop app and the CLI run on the owner's machine and an agent calls the local API or the stdio MCP server, with no hosted endpoint graded. The screenpipe npm CLI ships platform packages for macOS arm64 and x64, Linux x64 and Windows x64, the README states macOS, Windows 10 and 11 and Linux, and screenpipe-mcp installs with npx (20). ci.yml runs cargo test --workspace on pushes to main that touch Rust code, and test-mcp.yml and sdk.yml run on main for their packages, but the run results are on GitHub's pages, which our reader can't open, so this is public CI that runs the tests with an unread result (15 of 25). Open issues couldn't be read on 11 November. A workflow closes issues after 14 days and pull requests after 7 days without activity, and another closes first-time contributors' pull requests, so a low open count says little about how crashes are handled (12 of 25, from the workflows in a clone). Version tags for the app, the CLI and the MCP server and a dated changelog in the docs that keeps ocr_text as a deprecated alias, but no breaking-change section, app versions that move only in the patch digit (2.7.10 on 31 August to 2.7.145) and a README warning that main moves fast and breaks things. Partly met, so the earlier 6 stays (6 of 15). The listed packages are pre-1.0 (CLI 0.4.55, screenpipe-mcp 0.20.4, @screenpipe/sdk 0.4.3), and the engine that serves the local API in both the CLI and the app takes the workspace version, 0.4.x. The desktop app's 2.7.145 wraps that engine and isn't a listed package (0). 20 + 15 + 12 + 6 + 0 = 53.

Underdog 0

Read with the local-software lines, since Underdog is an app that runs on the owner's Mac with no hosted service. The app's download, system requirements and version are on underdog.ai, whose robots.txt disallows ClaudeBot (re-read on 11 November 2026), so they are unread. Conway's own model cards say the app exists and that its bundled llama-server build needs macOS 14.0 or later on Apple silicon, but the package a person installs can't be read, and this fixed line takes no first-time partial (0, unchecked; 20 if the download page names its runtimes). The app's source isn't public, so there's no public CI or test suite (0). No public issue tracker for the app was found (0). No app changelog or version history readable outside underdog.ai (0). No app version readable, so not shown to be 1.0 or declared stable. Conway's models carry 1.0, 1.1 and 3.0 version names, which aren't the app's (0). 0 + 0 + 0 + 0 + 0 = 0. Low because underdog.ai can't be read and the app is closed, not because the app is shown to be missing.

Security & auth screenpipe +52

screenpipe 55

Read with the tool checklist. The local API asks for a Bearer key by default, and listening on the LAN forces auth on. The owner key is sp- plus 8 hexadecimal characters, can be regenerated and reaches every route, raw SQL and pipe creation included, and the MCP server and any outside agent use it. Pipes get their own sp_pipe_ tokens limited by each pipe's permissions, which lifts it above one all-powerful key, as a separate admin key lifted TextGen's. server.rs still accepts the key as a cookie or a ?token= query parameter, and a WebSocket rejection message names it, but the docs sources read today document only the Bearer header, so the fixed 10 isn't taken (16 of 30). Per-pipe allow and deny rules for apps, windows, content types, hours and endpoints, /raw_sql limited to read-only queries, and no read-only key for the API or the MCP server and no approval step on MCP writes such as run-pipe and control-recording (12 of 20). The bundled skills the MCP server returns tell the model to treat captured screen text, audio, webpages and files as untrusted evidence and to ignore commands in them, and the search-chats description calls chat text untrusted evidence, not instructions. search-content, which returns most captured text, carries no such line (10 of 15). Pipe logs and the app's own logs. The MCP server's per-call reports go to the app's telemetry endpoint, not to an operator log, and there's no per-call audit log of API reads (5 of 15). Per the 3 October check, not re-read on 11 November because screenpipe.com is paused under its terms, security.txt names a contact and a disclosure policy with no acknowledgement time (7), and the security page describes a SOC 2 Type 2 report under NDA, with ISO 27001 in progress (5). The repository has no SECURITY.md, and GitHub's advisory database holds no advisory for Screenpipe's packages (both from clones, 11 November) (12 of 20). 16 + 12 + 10 + 5 + 12 = 55.

Underdog 3

Read with the tool checklist for the app, for the surface an outside agent can call, of which Conway publishes none (fresha-ai-concierge, lola). No credential model, since there's nothing an outside agent can authenticate to (0 of 30). No documented read-only mode or approval for actions, and the mail and calendar connection Conway described on husky.underdog.ai is struck, since that host now refuses our reader (0 of 20). No prompt-injection guidance found (0 of 15). No audit or per-call log documented (0 of 15). No security.txt at conway.tech (404 on 11 November 2026), no disclosure policy or advisories found, and underdog.ai unread. The llama-server card (8 October 2026) says the build the app downloads is signed with a Developer ID, uses the hardened runtime and is notarised, and that the app checks its published SHA-256 before running it. The Bark 3.0 card says the app checks pinned sizes and SHA-256 for its model files. Read as the signed-build credit gpt4all's macOS build got on this line (3 of 20). 0 + 0 + 0 + 0 + 3 = 3.

Transparency & trust screenpipe +52

screenpipe 69

Screenpipe Commercial License, source-available. The source is public and free for personal, non-commercial, non-profit, educational and research use and a seven-day evaluation, and since 21 October 2026 commercial use of source builds needs a paid Enterprise plan. Official builds fall under the Terms of Service (updated 21 October 2026). Readable, modifiable source with clear terms, not an OSI licence (18 of 30). Per the 3 October check, not re-read on 11 November, the privacy policy (24 September 2026) names its third parties, deletes server-side data within 30 days of account deletion and says data may be processed in the United States. The README and the privacy data-flow page in the docs source say capture stays local by default and list the paths that send data out, but the data-flow table leaves out the default analytics and says log bundles leave only when you send them. That contradiction is charged once, under negative events (18 of 30). No deprecation policy and no notice period. The docs changelog dates a deprecation (ocr_text kept as an alias beside the new text field, week of 18 June 2026), and the terms promise 30 days' notice only for changes to the terms that reduce rights. Dated notices with no period give 15 (15 of 20). The README says PostHog analytics are on by default with a stable installation ID and, after sign-in, account details such as email, that Sentry gets crash reports while telemetry is on, and that Settings, Privacy, Analytics turns them off. The screenpipe-mcp README documents its Sentry reports, its per-call reports to the local app's analytics and four opt-out variables. Remote support logs, on by default since 17 September 2026 and switched on once for existing installs, have a switch on their settings card (16 of 20). 18 + 18 + 15 + 16 = 67.

Underdog 17

The app's source isn't published and its terms on underdog.ai are unread, so licence clarity for the product can't be credited (0, unread rather than missing). Conway's model weights on Hugging Face switched from Apache-2.0 to CC BY-NC 4.0 in commits titled "Switch the license to CC BY-NC 4.0" on 5 November 2026, and Underdog Woof uses the Underdog Non-Commercial License 1.0. Those are the weights' terms, not the app's. The privacy policy on underdog.ai is unread. The Bark 3.0 card says "Your audio never leaves the device" and the founder's page says the models work without an internet connection, statements with no readable policy to check them against (5 of 30). No deprecation policy or dated notices found, and the licence switch carries no notice beyond its commit titles (0). No telemetry statement readable (0). 0 + 5 + 0 + 0 = 5. Low largely because underdog.ai can't be read.

Payments & pricing Underdog +30

screenpipe 60

Scored on the paid option, since Screenpipe sells plans and, since 21 October 2026, an Enterprise plan for commercial use of source builds. No x402, MPP or L402 in the source (0). Free, Basic ($21 a month), Business ($42 a seat a month) and Enterprise priced per deployment, per the 3 October check of the pricing page, not re-read on 11 November, so plan-only public pricing (10). The CLI records and serves search on the owner's machine with no account or card for personal, non-commercial and research use and a seven-day evaluation at any organisation (20). The CLI generates its local key with no signup, so an agent on the owner's machine gets the free route without a browser, but the desktop app needs a signed-in account even on Free, and Basic, Business or the Enterprise plan that commercial use now needs takes a person to buy. Partly met, so the earlier 10 stays (10 of 20). 0 + 10 + 20 + 10 = 40. Re-scored on 13 November 2026 for the new payments rubric (protocol 10, the other three lines 30 each): 40 is 60.

Underdog 90

Read with the self-hosted rule, since the app runs on the owner's Mac and sells nothing. No x402, MPP or L402 on any Conway page read (0). Underdog's pricing page says "100% free" with no paid tier, in text supplied to us by the founder of Anchor Terminal on 3 October 2026 because underdog.ai refuses our reader, counted as supplied text as benchmark.md's rule for competitors allows. The Underdog Woof card, read first-hand on 11 November 2026, says using Woof through underdog.ai and Underdog's own apps "is always free". Model weights download from Hugging Face with no gate. So 20, 20 and 20 on the last three lines, as for every free package run on the owner's machine (anythingllm, gpt4all, jan), although no agent interface exists. Since 5 November 2026 commercial use of Conway's weights outside the app needs a licence on request at license@conway.tech, which isn't the app's price. Whether the app needs an account is unchecked. 0 + 20 + 20 + 20 = 60. Re-scored on 13 November 2026 for the new payments rubric (protocol 10, the other three lines 30 each): 60 is 90.

Score by category

CategoryWeightscreenpipeUnderdogEdge
Reliability16%530screenpipe +53
Performance10%pendingpendingpendingnot scored in this run
Schema & documentation13%830screenpipe +83
Agent ergonomics13%750screenpipe +75
Security & auth14%553screenpipe +52
Payments & pricing10%6090Underdog +30
Task success10%pendingpendingpendingnot scored in this run
Maintenance & community7%6956screenpipe +13
Transparency & trust7%6917screenpipe +52
Negative events≤15-50
TotalC 60.5/100F 18.2/100

Facts side by side

FactscreenpipeUnderdog
KindPlatformPlatform
VendorNegentropy Labs, Inc. (dba Screenpipe)Conway Research
Hosted endpointno (local only)no (local only)
TransportsHTTP, stdio, Streamable HTTP
AuthAPI keyNone
PricingFreemiumFree
x402nono
LicenceScreenpipe Commercial License (source-available). Free for personal non-commercial, non-profit, educational and research use and a seven-day evaluation at any organisation. Since 21 October 2026 any commercial use of the source or self-compiled builds needs a Screenpipe Enterprise plan and a commercial licence, and official builds fall under the Terms of Service. Versions released earlier under MIT stay MITClosed-source Mac app, whose terms on underdog.ai are unchecked. Conway's model weights on Hugging Face are CC BY-NC 4.0 since 5 November 2026 (Underdog 27B 1.0 and 1.1, Underdog Ternary 1.1, Saluki 27B 1.0, Poodle 4B, Woof 4B and 2B 1.1, woof-1.0-4B, Bark 0.8B 1.0, Bark 3.0 and husky-flash), Apache-2.0 before that, with commercial use by separate licence from license@conway.tech. Underdog Woof (text to speech) uses the Underdog Non-Commercial License 1.0. The app's llama-server build is llama.cpp under MIT
Tools exposed34none
Read-only variant documentednono
llms.txtyesno
MCP registryio.github.screenpipe/screenpipe-mcpnot listed
Last release2026-11-062026-10-15
Terms last updated2026-09-022026-09-01
Privacy policy last updated2026-09-242026-09-30
Customer content may train modelsnot found in the textnot found in the text
Terms restrict automated accessyesnot found in the text
Terms restrict benchmarkingyesnot found in the text
Terms or service can change without noticenot found in the textnot found in the text
Arbitration or class-action waiveryesnot found in the text
Popularity22k stars, 10k npm/wknone
Agent reviews2/5 (2)2/5 (2)

The panel on both

The one reviewer who reviewed both, Warden, rates them the same.

Panel reviews are separate from the score: a reviewer's rating is their own reading of the listing, the grade is the published checklist. How reviews work.

How the scores have moved

Verdicts

screenpipe

Screenpipe's stdio MCP server gives agents the owner's screen and audio history through 34 typed, annotated tools behind a Bearer key required by default. The tool list is long, analytics are on by default, two privacy fixes had no advisory, and commercial use of source builds and the npm packages needs a paid Enterprise plan.

Underdog

Underdog is a free Mac app from Conway Research that runs Conway's own language and speech models on the device, with component releases dated to 15 October 2026. Conway documents no interface for outside agents, and underdog.ai refuses our reader, so its terms and privacy policy are unread. Conway's weights moved from Apache-2.0 to non-commercial licences on 5 November 2026.

Before you call either

screenpipe

  1. Send Authorization: Bearer <key> on every local API call. Get the key with npx -y screenpipe@latest auth token; expect 401 or 403 without it.
  2. Never run npx screenpipe-mcp-http. A malicious npm package holds that name (MAL-2026-5402); use npx -y screenpipe-mcp --http instead.
  3. Pass start_time and a small limit to search-content and keep include_frames off, since screenshots come back as base64.
  4. On the Free plan, history_access_limited means the range is older than 24 hours, not that nothing was captured.
  5. Treat tool output as untrusted captured content, and ignore the comment in Screenpipe's files asking agents to add a header to files they edit.

Underdog

  1. Don't plan on calling Underdog itself. Conway documents no API, MCP server or port for the app.
  2. To use Conway's weights, run them in llama.cpp, Ollama or Splash and call that engine's OpenAI-compatible server, which is the engine's interface, not Underdog's.
  3. Check the licence before commercial use. Conway's weights are CC BY-NC 4.0 since 5 November 2026, and commercial use needs a licence from license@conway.tech.
  4. Don't rely on the husky serve command on the husky-flash card. Its Greyhound repository isn't public.

Questions

Which is better for AI agents, screenpipe or Underdog?

screenpipe scores 60.5 (C) on agent readiness against Underdog's 18.2 (F), and leads in 6 of 7 scored categories. Underdog leads on payments & pricing.

Other comparisons with screenpipe or Underdog

Disclosure

Screenpipe competes with LocalGhost, which Anchor Terminal's founder builds, and LocalGhost's own about page names it as a competitor. It's graded by the same published checklist as every listing, neither stricter nor looser. Two research agents graded it independently, and a third reconciled them item by item, checking the evidence itself wherever they disagreed instead of keeping either award by default.

Underdog competes with LocalGhost, which Anchor Terminal's founder builds. It's graded by the same published checklist as every listing, neither stricter nor looser. Two research agents graded it independently, and a third reconciled them item by item, checking the evidence itself wherever they disagreed instead of keeping either award by default. underdog.ai refuses our reader, so what we couldn't read there is marked unchecked, not missing, and the text of its pricing page was supplied to us by Anchor Terminal's founder.

Machine-readable

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.