The AI says it is “done.” How do I know it is actually deliverable?

Still unverified

'Docs stay in GitHub' does not mean zero processing. The GitHub App needs read/write contents and pull-request permissions on granted repositories, and realtime collaboration state may persist temporarily for up to seven days. When a user invokes AI, document text and relevant comments go to xAI, whose API data may be retained for abuse monitoring for up to 30 days; explicitly shared AI conversations may be kept for up to 90 days. The current privacy page also discloses allowlisted, content-free analytics through PostHog Cloud EU: region- and consent-gated, with raw events retained up to 12 months and operational logs up to 30 days. These boundaries are more specific than older 'no analytics' summaries and deserve review before adoption.

Still unverified

Air is still a public preview. Available models, permissions, execution location, and cost depend on JetBrains AI, BYOK, or an external agent provider; an isolated worktree does not mean code or prompts never leave the device. Agents may edit files, run commands, or consume quota, so each diff, permission, and network path needs review. The live download page should take precedence over early platform plans; this review did not independently verify stability, privacy, speed, or agent compatibility.

Still unverified

Air is still a public preview. Available models, permissions, execution location, and cost depend on JetBrains AI, BYOK, or an external agent provider; an isolated worktree does not mean code or prompts never leave the device. Agents may edit files, run commands, or consume quota, so each diff, permission, and network path needs review. The live download page should take precedence over early platform plans; this review did not independently verify stability, privacy, speed, or agent compatibility.

Still unverified

'Code and prompts do not pass through the CodeAgentSwarm backend' does not mean the product has no data flows. The current policy says desktop usage events go to its backend; crash reports are forwarded to Sentry and stack traces may contain local file paths; optional sign-in stores OAuth profile data, tokens, IP address, and device information; and the Swarmi help assistant sends questions plus recent history to the backend and stores the exchange. The policy gives no fixed deletion period for analytics, errors, or support messages, and does not explain storage location, encryption, portability, or deletion for the complete local conversation history. The Turbo Mode guide also says Ask effectively behaves like Allow while confirmation skipping is on, so actions that must never auto-run need Deny. Conflicts from several agents editing one repository, guardrail coverage, and bypass paths were not independently tested.

Still unverified

The official pages use “90% fewer tokens,” “85% fewer tokens,” and an 85.7% estimate from one Reddit posting task; none should be treated as a general benchmark. The Chrome listing has no rating and the product is still early. Its privacy page discloses activeTab, scripting, debugger, nativeMessaging, clipboardWrite, and <all_urls> permissions, and lets a local agent evaluate short JavaScript expressions in the page; local execution is not the same as low risk. The repository uses Business Source License 1.1: free for personal use, with commercial use subject to the license.

Still unverified

The docs explicitly say full prompts, completions, tool arguments and results, and error context are captured by default; every hide flag for inputs, outputs, messages, images, embeddings, and model parameters defaults to false. Even maximum redaction keeps model names, token counts, timing, span types, and error messages, which can themselves contain sensitive material. The default environment guide requires Future AGI API and secret keys and points HTTP and gRPC collectors to its cloud. Custom collectors and bring-your-own TracerProvider are documented but were not exercised here. Uneed copy and the first launch image say MIT, while the current repository LICENSE, README badge, and NOTICE all say Apache-2.0. Uneed claims four-language full parity and 35+ frameworks, the images and README say 50+, the repository's own matrix contains many blanks, and cloud pricing says only 30+ libraries. The cloud free tier currently advertises 50 GB of tracing per month with 30-day retention and overage from USD 2 per GB. The data-region page says US is available to all plans and EU Frankfurt only to Enterprise, while the pricing card presents US and EU side by side. Security pages self-report SOC 2 Type II and ISO 27001 certification and HIPAA support through a BAA on Scale or Enterprise; they list AWS, ClickHouse Cloud, Cloudflare, Stripe, plus platform-wide HubSpot, PostHog, Sentry, and Intercom. No downloadable certificate, audit report, or penetration-test summary was public, so procurement must request them. The open-source instrumentation, Future AGI hosted platform, and its 'Self-host it' claim are three separate boundaries and should not be treated as interchangeable.

Still unverified

This is an early version, not a finished enterprise control plane. The team product is labeled early access. The Individual page still says v0.0.5 and 3.9 MB, while the download API pointed to v0.0.6 and the publisher screenshot also shows v0.0.6. No public changelog, installer SHA-256, reproducible build, or source code was found to map the binary. Feature definitions drift as well: the maker's Uneed copy says 47 hand-written rules while the current site says 25+; Individual says a pretrained classifier runs locally while the architecture labels it 'architecture shipped · model pending'; FAQ says 'no browser extension, no proxy' while Home and How it works include a local TLS proxy; and pages plus screenshots differ on coverage for Cursor, Claude Code, Copilot, Windsurf, and ChatGPT. The privacy boundary needs precision. Individual claims no account, backend, or network. Team mode may upload tool, category, severity, rule, timestamp, a redacted snippet up to 500 characters, and a correlatable hash. Privacy names Investigation Mode as the sole exception to no raw-prompt upload and retains investigation records for a contractual legal-hold period, while the architecture says the server stores an unreadable sealed payload. Public material does not explain whether that payload contains raw prompt text, when it is produced on the endpoint, who holds decryption keys, whether an employee can refuse or revoke it, what happens if notice fails, or the guessing risk of deterministic hashes over low-entropy secrets. Team events are queryable for 90 days, cold-stored for another 180, then retained as org aggregates. A public subprocessor page is still in preparation; the draft Privacy page names Supabase, ClickHouse Cloud, Clerk, Vercel, Fly.io, Resend, and Anthropic. The public site bundle also includes a PostHog page-view provider configured with memory persistence and session recording/autocapture disabled, while the current Privacy draft does not separately explain website analytics. Compliance and contracts are incomplete: Privacy and Terms are both April 22, 2026 drafts, and the binding terms and MSA are still being prepared with counsel. FAQ says SOC 2 Type II is in progress and HIPAA eligibility arrives with a future v1.5, while Pricing directly lists an Enterprise SOC 2 report, HIPAA eligibility, BYOK, and on-prem and also marks some items 'soon' in its comparison table. GDPR, signing/notarization, <50ms p95, 95% soft notifications, 99.9% SLA, the blog's three teams / 312 engineers, and enforcement of the no-performance-review covenant are team statements. No audit report, corpus, false-positive/negative results, notarization ticket, customer reference, or final contract was independently checked. Team and Business are annual per-seat subscriptions, and FAQ describes a $15K deposit for a 30-day paid POC. Before production procurement, confirm the available version, seat minimums, refunds, legal entity and address, lawful basis for employee monitoring, DPA, regions, subprocessors, keys, and deletion flow.

Still unverified

The official pages use “90% fewer tokens,” “85% fewer tokens,” and an 85.7% estimate from one Reddit posting task; none should be treated as a general benchmark. The Chrome listing has no rating and the product is still early. Its privacy page discloses activeTab, scripting, debugger, nativeMessaging, clipboardWrite, and <all_urls> permissions, and lets a local agent evaluate short JavaScript expressions in the page; local execution is not the same as low risk. The repository uses Business Source License 1.1: free for personal use, with commercial use subject to the license.