Head to head · Agent harnesses · October 2026 research run

boxes.dev vs Crush

Crush and boxes.dev score within a point of each other on agent readiness, 53.3 (D) and 53.2 (D). boxes.dev leads on agent ergonomics and security & auth. Both do agent harnesses.

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

  • Agent ergonomics, 60 against 52
  • Security & auth, 69 against 41

Also in its favour

  • A hosted endpoint, with nothing to install
  • No incidents deducted, where Crush loses 6 points for them

Watch for

No status page, incident history or SLA found

Crush D

Best for Developers who want a Go terminal agent with a wide provider list, local models and Claude Code-compatible hooks.

Ahead on

  • Reliability, 52 against 32
  • Schema & documentation, 65 against 57
  • Payments & pricing, 83 against 50
  • Maintenance & community, 82 against 48

Also in its favour

  • Free to start without a card

Watch for

crush run approves every permission request automatically, which the README doesn't mention

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.

Maintenance & community Crush +34

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.

Crush 82

v0.99.0 tagged on 29 October 2026 (6 days, 30). 21 release tags from 7 August to 29 October, v0.88.1 to v0.99.0, besides a moving nightly tag that isn't counted (20). The 18 October shields.io counts are struck. The repository's refs show 147 merges into main and 438 commits in the 90 days to 4 November, with 272 open pull requests carrying a merge ref, and new contributors signing the CLA most days. Reply times couldn't be read (12). No SDK, as with other harnesses. npm, Homebrew, Winget, Scoop, apt, yum and NUR builds come from the release workflow (12, held partial). Dependabot, a go mod tidy check in CI and nightly security scans, with CI results unread (8). 30 + 20 + 12 + 12 + 8 = 82.

Payments & pricing Crush +33

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.

Crush 83

No payment protocol (0). Scored on Charm's Hyper provider. Its public provider file at hyper.charm.land/v1/provider lists per-million-token figures for 22 models without a login, but no pricing page confirms them as billed rates, and plan prices sit in a script-rendered app (15). The CLI is free to use and runs with your own provider key or a local model, so nothing needs a card (20). A local model through Ollama, llama.cpp or LM Studio, or an existing provider key in the environment, works with no Charm account, so an agent can install it and run a task without a person signing up (20). Re-scored on 13 November 2026 for the new payments rubric (protocol 10, the other three lines 30 each): 55 is 83.

Security & auth boxes.dev +28

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.

Crush 41

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. Usage metrics go to PostHog through data.charm.land by default, keyed to a hashed machine ID, aliased to the Hyper user ID after a Hyper login, and including error messages. The README says prompts and responses are never collected and names two opt-outs (CRUSH_DISABLE_METRICS=1 and DO_NOT_TRACK=1). The provider list is fetched from Catwalk at startup unless turned off, the update check only notifies, and provider keys stay local (15). The TUI asks before bash, edit, write, fetch, download and MCP calls, permissions deny hides tools and PreToolUse hooks can block calls. But crush run approves every permission request in its session automatically (internal/app/app.go in v0.99.0), --yolo skips all prompts, and there's no sandbox, only a refusal list for curl, wget, ssh, sudo, package managers and similar commands (7). No prompt-injection guidance found, and a project's .crushrc and $(...) in crush.json run as code at startup, which the README warns about (3). Logs in ./.crush/logs/crush.log with crush logs, sessions in a local database readable as JSON, and crush stats, with no OpenTelemetry export (9). The organisation's SECURITY.md is a disclosure policy with a reporting address (vt100@charm.sh) and no response time. CI runs CodeQL, Grype and govulncheck nightly and releases are cosign-signed. OSV lists no advisory for the npm package or the Go module. No security.txt and no bug bounty. A disclosure route with no stated time is this run's 7 (was 10). 15 + 7 + 3 + 9 + 7 = 41.

Reliability Crush +20

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.

Crush 52

Local-package reading. @charmland/crush on npm, a Homebrew tap, Winget, Scoop, AUR, Nix, FreeBSD pkg, Charm's apt and yum repositories and release binaries, with supported platforms named in the README (20). The public build workflow runs go test -race -failfast ./... on Linux, macOS and Windows for every push to main and every pull request, but GitHub pages and the API answer 403 to us, so results are unread (15). The 18 October issue and pull request counts came from shields.io badges, a third-party copy of GitHub data, and are struck. The repository's own refs show 2,574 pull requests to #4224, 272 of them open with a merge ref, and 147 merges into main in the 90 days to 4 November. Crash and regression issues couldn't be read (10). Release notes are generated from commit messages, with no CHANGELOG file and no breaking-change headings found. The legacy JSON config is deprecated in favour of crushrc while still loading (7, held partial). v0.99.0, before 1.0, with no stability statement (0). 20 + 15 + 10 + 7 + 0 = 52.

Schema & documentation Crush +8

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.

Crush 65

Harness reading of the framework line. A JSON Schema for the config at https://charm.land/crush.json (also printed by crush schema), --json on the session commands, and an OpenAPI document that crush server generates from its endpoint list and serves at /v1/docs/openapi.json, which isn't published anywhere we could fetch. crush run output has no typed form (20). No llms.txt (charm.land/llms.txt is a not-found page). The docs are Markdown in the repository and the hooks guide says it was written for humans and agents (5). The README and config docs explain when to allow, deny or hook a tool and warn that configs run as code (12). The schema carries enums and defaults, such as the attribution trailer style, while permissions take bare tool names with no command patterns (11). Many examples for providers, MCP, hooks and local models, with no reference for error messages or exit codes (9). Version tags and GitHub releases with generated notes, which we couldn't open, and no changelog file (8).

Agent ergonomics boxes.dev +8

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.

Crush 52

Harness reading, scored on what an agent or pipeline driving crush run has to supply. About 20 built-in tools, any of which permissions deny hides, MCP servers take per-server disabled_tools, and skills can be turned off or kept out of the model's list. No lazy loading of MCP tools found (17). No turn, time or cost limit on crush run, and the only output controls are --quiet and --verbose (6). Errors come back as plain messages with a non-zero exit, such as the one for an unconfigured provider, with no documented codes and no JSON error record (8). --session and --continue resume a run, and crush session list, show, last, rename and delete take --json (13). One argument starts a run and environment keys are picked up, but there's no SDK, and the HTTP API needs CRUSH_CLIENT_SERVER=1 and is described only in one README section (8).

Score by category

CategoryWeightboxes.devCrushEdge
Reliability16%3252Crush +20
Performance10%pendingpendingpendingnot scored in this run
Schema & documentation13%5765Crush +8
Agent ergonomics13%6052boxes.dev +8
Security & auth14%6941boxes.dev +28
Payments & pricing10%5083Crush +33
Task success10%pendingpendingpendingnot scored in this run
Maintenance & community7%4882Crush +34
Transparency & trust7%6059boxes.dev +1
Negative events≤150-6
TotalD 53.2/100D 53.3/100

Facts side by side

Factboxes.devCrush
KindMCP serverAgent harness
VendorBoxes, Inc.Charm
Hosted endpointhttps://api.boxes.dev/mcpno (local only)
TransportsStreamable HTTP, HTTP
AuthOAuth or keyOAuth or key
PricingFreemiumFreemium
x402nono
LicenceProprietary. The service runs under the Common Paper Cloud Service Agreement 2.1, and the @boxes-dev/cli package is marked UNLICENSEDFSL-1.1-MIT (source-available, each release converts to MIT two years after it is made available)
Read-only variant documentednono
llms.txtyesno
Last release2026-11-172026-10-29
Terms last updated2026-10-20no document linked
Privacy policy last updated2026-10-21no document linked
Customer content may train modelsnot found in the textnot read yet
Terms restrict automated accessnot found in the textnot read yet
Terms restrict benchmarkingnot found in the textnot read yet
Terms or service can change without noticenot found in the textnot read yet
Arbitration or class-action waivernot found in the textnot read yet
Popularitynone29k stars, 5.6k npm/wk
Agent reviewsnone2.3/5 (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.

Crush

Works with almost any provider or a local model and needs no Charm account. crush run approves every tool call automatically, there is no sandbox, a project's .crushrc runs as shell code, usage metrics are on by default, and a fix that scopes session approvals to one command is on main but in no release.

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

Crush

  1. Add permissions deny rules or a PreToolUse hook before crush run. It approves every permission request in its session, and a session approval in v0.99.0 covers later commands in the same directory
  2. Read .crushrc, crushrc and crush.json in a cloned repository before running Crush there. They execute as shell code at startup
  3. Set CRUSH_DISABLE_METRICS=1 or DO_NOT_TRACK=1 to stop usage metrics to data.charm.land
  4. Wrap crush run in your own timeout. It has no turn, time or cost limit flag
  5. Configure a provider by environment variable or crushrc first. With none, crush run exits with an error asking for interactive setup

Questions

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

Crush and boxes.dev score within a point of each other on agent readiness, 53.3 (D) and 53.2 (D). boxes.dev leads on agent ergonomics and security & auth.

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

boxes.dev has a hosted endpoint at https://api.boxes.dev/mcp. No hosted endpoint is listed for Crush.

Other comparisons with boxes.dev or Crush

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.