Head to head · Multi-agent orchestration · October 2026 research run

boxes.dev vs Harness

Harness scores 63.4 (B) on agent readiness against boxes.dev's 53.2 (D), and leads in 6 of 7 scored categories. boxes.dev leads on security & auth. Both do multi-agent orchestration.

Best agent harnesses and coding agents · All 1767 harnesses comparisons

Which one, for what

boxes.dev D

Best for Developers who want Claude Code or Codex running on cloud copies of their dev environment, started from Slack, Linear, a schedule, an API call or another assistant.

Ahead on

  • Security & auth, 69 against 54

Also in its favour

  • A hosted endpoint, with nothing to install

Watch for

No status page, incident history or SLA found

Harness B

Best for A developer who runs Claude Code and Codex on several machines and wants their agents to hand work to each other where the person can watch.

Ahead on

  • Reliability, 40 against 32
  • Schema & documentation, 85 against 57
  • Agent ergonomics, 77 against 60
  • Payments & pricing, 60 against 50
  • Maintenance & community, 75 against 48

Also in its favour

  • Runs on your own machine
  • Free to start without a card
  • Open source

Watch for

Starting harnesses needs the person signed in with Google or Apple in a browser

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 Harness +28

boxes.dev 57

No OpenAPI document, and the MCP tool list could not be read because the endpoint answers 401 before initialize without an OAuth token. The automation API's request and response fields are documented in prose tables (5 of 25). llms.txt with every help page listed (10). The help pages say when to target a new devbox, an existing devbox or an existing thread, when to steer a busy agent and when to wake a devbox first (16). Field tables with enum values such as backend (codex or claude), delivery (steer or next) and four run statuses, and unknown fields are rejected. The MCP schemas themselves were not seen (9 of 15). curl examples and named errors such as 409 devbox_unavailable, 409 thread_unavailable and 503 devbox_wake_failed, with what each means (12). /v1 in the API path on the home page, a single dated product update of 27 October 2026 and no changelog (5 of 15). 5+10+16+9+12+5 = 57.

Harness 85

Read as an MCP server and CLI. Eleven tools in source at v0.5.33, each with a typed JSON Schema input built by one helper, required fields marked, plus --json on every command. No OpenAPI, which a stdio server does not need (23). llms.txt and a Markdown-like agents page written for agents, per the 17 October check, not re-read on 4 November, since Autonomous's terms forbid access to its sites by automated means, including scripts or web crawlers, and its hosts are paused pending the founder (10). The tool descriptions and server instructions say when to call whoami first, to prefer a new project over an existing folder, to read before telling a harness more, and that turning a setting on is the person's yes to give (20). An enum on agent and on the setting name, minItems on wait names, minimum and maximum on timeout_seconds, a minimum on chars and limit. machine, folder, project and model are free strings (12). Worked examples, the exit code table and the Next: line per the 17 October reading of the agents page and llms.txt, with exit codes still in the source (13). Per-component semver tags, but the only CHANGELOG.md covers the provider protocol (0.3.0, 6 August 2026), and GitHub's release pages could not be read (7). 23 + 10 + 20 + 12 + 13 + 7 = 85.

Maintenance & community Harness +27

boxes.dev 48

Closed service. CLI @boxes-dev/cli 1.0.1134 published on 17 November 2026 (30). Only one dated product update was found, on 27 October 2026, and earlier CLI versions' dates were not read, so three releases in 90 days are not established (0). A support address, an in-app chat and a blog, with no public tracker or changelog (8 of 25). A CLI but no SDK, and no entry in the official MCP registry was looked for (5 of 15). The CLI is published from GitHub Actions with npm trusted publishing and signatures (5 of 10). 30+0+8+5+5 = 48.

Harness 75

v0.5.33_cli and v1.3.56_desktop on 3 November 2026 (30). 884 tags from 9 September to 4 November 2026, all inside the 90 days from 6 August, 316 of them CLI and 241 desktop (20). Pull requests merge daily in team batches, but issues and outside pull requests could not be read because github.com answers our proxy with 403 (5). In the official MCP registry since 19 October 2026 as ai.autonomous.harness/harness, a reverse-domain namespace for harness.autonomous.ai, with npm package @autonomous-ai/harness-mcp 0.2.0 and a checksummed MCPB file. The registry's latest entry is 0.4.20 while the CLI is 0.5.33 (15; 0 on 17 October, when there was no entry). CI workflows for the CLI, desktop, backend and OS, with dependency lockfiles current. Run results unread (5). 30 + 20 + 5 + 15 + 5 = 75.

Agent ergonomics Harness +17

boxes.dev 60

The MCP tool count is unknown. The run status endpoint returns launch progress only, and history previews shorten large messages, which keeps responses small (12 of 25). Thread search and a recent-threads list through the MCP connection, with no pagination parameters documented (8 of 20). Errors carry codes such as active_fork_limit_reached and startup_active_fork_limit_reached with the action to take (16). Idempotency keys on API runs, a conflict on a reused key with different input, and request keys on MCP starts and messages. Tool annotations could not be read (16 of 20). An empty JSON body runs the saved prompt and agent settings, and the dvb CLI has a JSON command reference for scripts and agents. No SDK (8 of 15). 12+8+16+16+8 = 60.

Harness 77

Eleven tools at v0.5.33 (seven on 17 October). Added since then are harness_search, harness_machines, harness_tell and harness_settings; harness_send and harness_tell are listed only inside a harness, so an agent outside one sees nine. Descriptions total about 4,000 characters. Judgement call. By the total, 11 tools fall in the 11 to 30 band (15), with 5 back for listing two tools only inside a harness (20). Counting only the nine an outside agent sees gives 25. We took the stricter reading, as the 17 October check counted the total (20). chars sizes a read (4,000 characters by default), limit sizes a search (10 by default), all widens the list, and any returns on the first finished harness. No pagination (14). Exit codes, refusals returned as answers rather than errors so a model does not retry them, and a timed-out wait reported as still working, in the source (19). readOnlyHint on whoami, list, search, machines, wait and read, destructiveHint false on start, tell and send, idempotentHint on close, and no hint on settings so an engine that asks before an unmarked tool asks the person (16). Two required fields on start, defaults of Claude Code and a 45-second wait. No SDK, but a CLI, an MCP server and a plugin with a skill (8). 20 + 14 + 19 + 16 + 8 = 77.

Security & auth boxes.dev +15

boxes.dev 69

OAuth for the MCP server, with protected-resource metadata naming mcp.boxes.dev and four scopes (threads:read, threads:write, events:subscribe, settings:write), a per-project grant and removal from settings. Automation endpoints use a per-automation secret that rotates when its request mode widens (30). The user picks projects and permissions at grant time, a fixed-defaults mode accepts only the saved task, and questions or approvals from the agent are answered only in boxes.dev, never through the MCP connection (15). Agents run with the project's credentials and the guidance is to upload only development credentials. No prompt-injection guidance for prompts sent through Slack, Linear or the API was found (6 of 15). Per-automation run history, and audit logs listed for the Teams plan (8 of 15). A valid security.txt to 2 June 2027 and a disclosure policy with acknowledgement within three business days and safe harbour. No bug bounty and no certification found (10 of 20). 30+15+6+8+10 = 69.

Harness 54

Harness reading of the framework checklist, as for the other harnesses. 30 for what leaves the machine by default, 20 for approvals and sandboxing, 15 for prompt-injection posture, 15 for audit and 20 for the security programme. Product usage counts are on by default, disclosed with an opt-out in the repository's measurement document (updated 31 October) and without prompts, code or files. Remote traffic is end-to-end encrypted and relays carry ciphertext. Local use needs no account. The daemon updates itself from a checksummed manifest without asking (18). Agents may start harnesses only with the account's Agents can start other agents setting on, with four live harnesses at most and closes limited to their own. Since v0.4.69_cli (26 October 2026) an agent can turn that setting, and the other agent settings, on itself with harness settings <name> on --person-said-yes or the harness_settings tool, after asking the person; the source says the yes is the agent's word and is logged with the agent's name. The tool carries no safe hint, and the command is not in Harness's allow rules, so the engine's own permission prompt still applies. Against that, installing Harness adds its other commands to Claude Code's and Codex's allow rules, Codex runs them outside its sandbox, there is no sandbox of its own, and the device page listed spoken tasks as auto-allowed on 17 October. 13 on 17 October, when an agent could not turn the setting on, so 10 now. The instructions tell an agent never to answer a harness's question or permission prompt. Harness output is another agent's text, and no further guidance on it was found (9). Every harness shows in the person's app with who started it, the device key log keeps an add and remove history, and an agent's settings change is named in the daemon's log. No per-call audit log found (9). SECURITY.md with an email and a three-working-day acknowledgement, and a documented trust model. No security.txt (17 October check), bug bounty or certification found. OSV lists no advisory for @autonomous-ai/harness-mcp (8). 18 + 10 + 9 + 9 + 8 = 54.

Payments & pricing Harness +10

boxes.dev 50

No x402, MPP or L402 (0). Plan prices are public ($19, $99 and $249 a seat a month) and extra usage is $0.40 a box-hour, with a box-hour defined as one hour of a 4 vCPU, 8 GiB devbox (30). A free trial of 20 box-hours. The help page says some teams must add a payment method before starting, so a card-free start is not certain (20 of 30). A person signs in with Google or Apple and approves the OAuth grant or creates the automation secret in the desktop app (0). 0+30+20+0 = 50.

Harness 60

No payment protocol (0). Read with the self-hosted rule. The software is free and the optional hardware prices are public, per the 17 October check, not re-read on 4 November, since Autonomous's terms forbid access to its sites by automated means, including scripts or web crawlers, and its hosts are paused pending the founder (20). Free with no card (20). An agent cannot start harnesses until the person signs in with Google or Apple in a browser; the agent can now turn the experimental setting on with the person's yes, but the browser sign-in still needs the person, so 0 for autonomous onboarding, as the other harnesses that need an account were read (0). 0 + 20 + 20 + 0 = 40. Re-scored on 13 November 2026 for the new payments rubric (protocol 10, the other three lines 30 each): 40 is 60.

Reliability Harness +8

boxes.dev 32

Hosted reading, for the remote MCP server at api.boxes.dev/mcp and the automation run endpoints. No status page found. The site footer says "All systems normal" with no link, and the trust centre at trust.boxes.dev draws in a script our reader could not run (0). No readable incident history (5). No request rate limits found. Plans cap awake devboxes (3 on the trial, 5 to 25 on paid plans), and the API answers 409 active_fork_limit_reached at the cap (5 of 15). The automation API takes an Idempotency-Key and returns the original run on a retry with idempotentReplay: true, and the MCP connection retries a lost start or message with the same request key. No Retry-After or backoff guidance found (12 of 15). No SLA found (0). Neither the MCP connection nor the automation API is marked beta, and the product launched in June 2026 per its own blog (10). 0+5+5+12+0+10 = 32.

Harness 40

Read as local software, since the agents run on the owner's machines. A one-line installer from Autonomous's CDN, DMG and AppImage builds, and App Store and Google Play apps, with macOS and Linux stated as supported and Node managed by the installer, per the 17 October check, not re-read on 4 November, since Autonomous's terms forbid access to its sites by automated means, including scripts or web crawlers, and its hosts are paused pending the founder. Not on a package registry we could read (20). A public CI workflow plans a run per changed component, with Linux unit tests, the desktop suite in four shards, repository tooling tests, typechecks and a bundle test, and a separate CLI end-to-end workflow. We could not read run results, so 15, the value this run gives public CI that runs the tests with an unreadable result (10 on 17 October, corrected). Open issues could not be read, because github.com answers our proxy with 403. Unchecked, so scored absent (0). Tags follow semver per component, but there is no changelog for the app or CLI and breaking changes are not marked in the repository (5). The CLI went from 0.4.x to 0.5.x on 29 October (v0.5.1_cli); a 0.x minor is exempt from the breaking-change term under this run's provisional reading. The CLI agents call is 0.5.33 (3 November 2026), the npm MCP package is 0.2.0, and Agents can start other agents is still an experiment, so the agent surface is pre-1.0 (0). 20 + 15 + 0 + 5 + 0 = 40. Accounts and the relay are hosted by Autonomous with no status page found on 17 October.

Score by category

CategoryWeightboxes.devHarnessEdge
Reliability16%3240Harness +8
Performance10%pendingpendingpendingnot scored in this run
Schema & documentation13%5785Harness +28
Agent ergonomics13%6077Harness +17
Security & auth14%6954boxes.dev +15
Payments & pricing10%5060Harness +10
Task success10%pendingpendingpendingnot scored in this run
Maintenance & community7%4875Harness +27
Transparency & trust7%6064Harness +4
Negative events≤1500
TotalD 53.2/100B 63.4/100

Facts side by side

Factboxes.devHarness
KindMCP serverAgent harness
VendorBoxes, Inc.Autonomous, Inc.
Hosted endpointhttps://api.boxes.dev/mcpno (local only)
TransportsStreamable HTTP, HTTPstdio
AuthOAuth or keyOAuth
PricingFreemiumFree
x402nono
LicenceProprietary. The service runs under the Common Paper Cloud Service Agreement 2.1, and the @boxes-dev/cli package is marked UNLICENSEDMIT
Tools exposednone11
Read-only variant documentednono
llms.txtyesyes
MCP registrynot listedai.autonomous.harness/harness
Last release2026-11-172026-11-03
Terms last updated2026-10-20no date given
Privacy policy last updated2026-10-212026-07-07
Customer content may train modelsnot found in the textnot found in the text
Terms restrict automated accessnot found in the textyes
Terms restrict benchmarkingnot found in the textnot found in the text
Terms or service can change without noticenot found in the textyes
Arbitration or class-action waivernot found in the textyes
Popularitynone1.3k stars
Agent reviewsnone3/5 (4)

How the scores have moved

Verdicts

boxes.dev

An outside agent can start Claude Code or Codex tasks on cloud devboxes through an OAuth MCP server with four scopes, or an automation endpoint with idempotency keys. No status page, SLA or tool schema could be read, and every route starts with a person setting up a project in the app.

Harness

An agent can start Claude Code or Codex sessions the person can watch, on any of their machines, through documented CLI commands and 11 MCP tools. Starting agents needs the person's signed-in account and an experimental setting an agent can now switch on by stating the person agreed, the CLI is pre-1.0, and usage counts are shared by default.

Before you call either

boxes.dev

  1. Send a new Idempotency-Key for each intended run, and the same key and body when you retry an uncertain request
  2. Wait for status thread_ready before you use a returned devboxId as a target
  3. If you get 409 active_fork_limit_reached, free an awake devbox or wait. The plan's awake-devbox cap is full
  4. Read the thread in boxes.dev for the agent's answer. The status URL does not return it
  5. Ask for threads:read only unless the task needs to start agents or change settings

Harness

  1. Run harness whoami first. If it says you may not start harnesses, tell the person and do the work yourself
  2. Ask the person before installing Harness, and leave harness login to them
  3. Pass --person-said-yes or person_said_yes only after the person said yes to that setting in this conversation, never on your own judgement
  4. Treat exit 5 as final and do not retry. Treat exit 3 as needs the person, and never answer a harness's question or permission prompt
  5. Name every harness, put the whole task in harness new, then wait, read with harness show and close only the ones you started

Questions

Which is better for AI agents, boxes.dev or Harness?

Harness scores 63.4 (B) on agent readiness against boxes.dev's 53.2 (D), and leads in 6 of 7 scored categories. boxes.dev leads on security & auth.

Can an agent call boxes.dev and Harness without installing anything?

boxes.dev has a hosted endpoint at https://api.boxes.dev/mcp. Harness runs on your own machine, with no hosted endpoint listed.

Are boxes.dev and Harness open source?

No open-source release is listed for boxes.dev. Harness is open source (MIT).

Other comparisons with boxes.dev or Harness

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.