Point Swarmat your code.Free.

Yes. Install it into Claude Code, Codex, or Cursor with `claude mcp add swarm npx @swarmsec/swarm-cli`, then ask your agent to run a Swarm scan. It reads your source and maps your live footprint for free, in minutes, with no account and no card. It runs a handful of Swarm's 50+ specialists on your own code and domain. The full engagement, run by an independent third party against your deployed app, is the paid step; pricing lives at https://swarmsec.ai/pricing.

No. The CLI ships as an MCP server that uses your host agent's model, so Claude Code, Codex, and Cursor drive it with the subscription or key you already have. There is no new key to set. For headless and CI use, `claude -p "run a swarm scan" --output-format json` runs the same scan against a provider key that is already a pipeline secret.

A one-shot code review reads your source in a single pass and stops there. Swarm is agentic and multi-specialist, and it looks at both your source AND your live attack surface. Specialist agents read the code the way a senior pen tester would (hardcoded secrets, missing authorization, known-vulnerable dependencies, stack-specific bugs across Node, Python, Go, Ruby, .NET, and PHP) and passive recon maps what the internet can already see about your domain: exposed services and admin panels, spoofable SPF and DMARC, dangling subdomains, public leaks. Candidate findings come back machine-actionable so your own AI session can start fixing them.

No, and that boundary is the point. A scan you run on yourself flags what looks wrong; it cannot prove the bug is exploitable, and an auditor will not accept a test you ran on your own code. Proving exploitability needs a deployed app plus authorization, which is the full engagement run by an independent third party. That engagement actively exploits and validates each finding, ships a working proof-of-concept with every one, chains low-severity findings into critical attack paths, and hands you an audit trail your prospect's security team can replay. That is third-party validation (SOC 2 CC4 / CC7), which is what your auditor requires.

For most SaaS engagements driven by SOC 2 Type 2 readiness, yes. That is the core use case. As a human pen test alternative and ethical hacking service, Swarm replaces the standard annual engagement for the majority of SaaS security programs. A human pen test firm typically charges $15,000 to $50,000 per engagement, takes two to four weeks, and delivers a PDF whose methodology lives in the consultant's head. Swarm runs in roughly two hours per engagement and ships a structured report plus the full audit trail of every specialist action: receipted, filterable, traceable from any finding back to the request that surfaced it.

Swarm is priced per engagement, not as a subscription or retainer, so you test as often as you ship. Customers run for SOC 2 Type 2 or ISO 27001 audit prep, and re-run as they ship for post-incident validation, new-feature security review, or security-questionnaire responses. Re-testing a fix is just another run, the close-the-loop validation that human firms charge separately for. Pricing detail lives at https://swarmsec.ai/pricing.

What Swarm replaces well: standard SaaS pen test engagements, especially the recurring annual or semi-annual ones, and especially when an external auditor is the deal-closing reviewer. The combination of an evidence-driven orchestrator dispatching 50+ specialists, the live activity feed, and the full forensic audit trail typically gives auditors more methodology transparency than a human-firm PDF.

What Swarm does not replace: bespoke red team assessment engagements with sophisticated social engineering, on-premise hardware testing, or multi-month engagements scoped to a specific advanced-persistent-threat hypothesis. For those, hire a senior firm. For the SOC 2 pen test you run every year, run Swarm and put the savings into remediation.

No. Automated scanners match known signatures against a checklist. Swarm specialists reason. They build a model of how your application works, form hypotheses, and test them adaptively. The result is findings scanners cannot produce: logic flaws, chained exploits, and authentication bypasses that do not appear in any CVE database. The CVE library augments this; specialists consult it for known issues. But the core engine is reasoning, not signature matching.

Yes. After every engagement, the swarm reviews what just happened and rewrites six knowledge bases that feed dispatch decisions. The mechanism is Anthropic Dreaming (beta), a research capability that lets agents reflect on completed work and update their own context. Swarm runs it against six surfaces: environment signals (stack detection patterns the orchestrator uses to choose specialists), per-specialist lessons learned, orchestrator dispatch heuristics, the CVE curation that decides which disclosures matter for offensive work, the compromise-pattern catalogue refined against new incident reports, and a false-positive refinement loop that updates the environment model whenever a finding gets rejected on review.

The practical effect lands at the platform level: across all engagements, the orchestrator routes specialists faster, the reviewer rejects fewer false positives, and the chain analyst recognizes exploit-chain shapes it has seen before. None of this requires a release on our side; the knowledge bases compound passively between runs.

Dreaming runs only on completed engagements. Abstracted lessons (CVE relevance, exploit-chain shapes, dispatch heuristics) inform the platform; per-customer signals stay scoped to your organization at the same data-model layer that enforces engagement ownership, so cross-tenant leakage is structurally impossible.

Yes. The deliverable is designed for SOC 2 Type 2 review and accepted as a compliance pen test deliverable by SOC 2 auditors. The SOC 2 Type 2 pen test report includes executive summary, individual findings with CVSS scores, exploit chain analysis, and validated proof-of-concept for every finding. The OWASP audit coverage maps every finding to its OWASP category (OWASP Top 10 testing plus OWASP API, LLM, and Agentic Applications Top 10) so your security questionnaire answers write themselves. The full audit trail (every specialist action receipted, filterable by specialist, traceable from any finding back to the request that surfaced it) gives your external auditor forensic-level transparency into methodology. A dedicated read-only Auditor role lets your compliance professional access the dashboard, report, and full audit trail directly.

Yes. Specialists operate within a customer-approved scope before testing begins. No destructive operations are taken without explicit per-action approval. Rate limits are enforced. Every request is logged and exported in the audit trail. Out-of-scope hosts are rejected at the tool layer before any HTTP call leaves the orchestrator.

A traditional pen test delivers a PDF and a verbal debrief; the methodology lives in the consultant's head. Swarm logs every move every specialist makes (every HTTP request, every source grep, every file read, every finding submission, every exploit chain composition) and streams it to your dashboard as the engagement runs. Hand the full record to your SOC 2 auditor afterward. They filter by specialist, pivot the dataset, and trace any finding in the report back to the exact tool call that surfaced it. Methodology that proves itself, not a summary that asks to be trusted.

Swarm specialists work against any modern web stack: Node, Python, Go, Ruby, Elixir, JVM, .NET, PHP. Coverage extends across every major identity provider too: Clerk, Auth0, Okta, Stytch, Cognito, Firebase, Supabase, and custom IDPs. The orchestrator fingerprints your stack during recon and dispatches the appropriate specialists automatically. AI / LLM and MCP server testing kicks in when those surfaces are detected, so you do not configure specialist-by-specialist; the swarm reads the application and routes work accordingly.

Yes. Mint a per-engagement Model Context Protocol token from the dashboard, plug it into Claude Code, Cursor, or any MCP-compatible client, and your team's editor surfaces Swarm findings, the source files the specialists already pulled, and a finding-status update tool in one place. Seven curated tools cover read access to findings and repositories plus the single write path of marking a finding remediated. Tokens are scoped to a single engagement and revoked with one click; nothing in the token can touch another engagement.

The intended workflow: an engagement closes, your engineers open the report inside Claude Code, fetch each finding's full evidence inline, write the fix against the source the specialists already read, and mark the finding remediated from the editor. The 30-day re-test window then validates the fix without a separate scoping call.

The service token is stamped with a developer role: reads plus finding-status updates only. It cannot run engagements, edit scope, change billing, or reach another organization's data.

Swarm is priced per engagement, with no subscription or retainer, so you test as often as you ship. Human pen test firms typically charge $15,000 to $50,000 per engagement and take two to four weeks. Every Swarm run includes the full deliverable: structured report, audit trail of every specialist action, and a validated proof-of-concept for every finding. Re-testing a fix is just another run. See current pricing at https://swarmsec.ai/pricing.