All the capabilities, by workspace.

See. One screen shows the trust grade, the alerts that matter, the traffic the gateway handled and where compliance stands. A count that cannot be read shows a dash, never a zero.

Govern. Findings from testing, scanning, intelligence, audit and compliance land in one issue with an owner and a due date. A risk acceptance has to expire, so work cannot quietly disappear.

See. Measured insights such as overdue fixes, untested agents and spend spikes, each with its evidence and the rule that fired. An AI brief ranks them and must cite them.

Govern. Five pillars, graded A+ to F, with what pulls the score down and which evidence is missing. A grade resting on thin evidence is labelled provisional.

See. A live health check lists each service of the platform with its current status, so IT can confirm the workspace itself is running.

See. An interactive map of the AI supply chain: model providers, agents and the classified data around them, with live status on each node. Where a list cannot be read, it says so.

See. Nine lookups and nine confirm-first actions run under your own permissions. When an answer is an action, it arrives as a card you approve, never something the model runs itself.

Govern. Published policies show their version and owner. People accept each one by version, and a new version asks again, leaving the record an auditor will ask for.

Control. One screen shows request volume against blocked requests, policy matches, provider health and any active kill switch. A figure that cannot be read shows a dash, never a zero.

Control. Connect 31 provider families, including self-hosted models. Your keys are sealed at rest and never shown again, and provider health is scored from the traffic that actually passed.

Control. Each policy row says whether the control is really in force and why not, if it is not. A count shows requests where a control could not run.

Control. Set which people and contexts may use which models, with per-request and daily allowances. Switch any model off in one place; retired models are never offered.

See. Each request that reaches the gateway is stored with the person, the model that answered, tokens, cost and whether a control stopped it. Per-request decisions can be read back through the API.

Prove. Try a prompt, or build and test a whole agent, through the same controls production traffic meets. Then register it and send it for approval.

Control. Halt AI for the whole workspace, one provider, one model or one person, with a written reason on record. Provider and model switches catch requests that name them.

See. Discover AI tools from gateway traffic, 13 collector types and a browser sensor. One decision becomes a standing rule, and the product says where a block really takes effect.

Control. Add your own identifiers, such as a policy number format, and try them on sample text. Plant canary tripwires in prompts, documents or database rows: a leak opens a critical incident.

Govern. One inventory of data assets with their classification, retention and recent consent. Retention flags, archives or anonymises on schedule and never hard-deletes.

Control. When a person withdraws consent for AI use, the gateway refuses their next request and logs it. It covers the inference-context purpose for signed-in users, and records a gap when the check cannot run.

Control. A control view of the MCP Firewall rules: allow, monitor or block each tool an agent can call. It is a rule list, not a discovered inventory of tools.

See. Learned from real requests, nothing added by hand: who reached each server and which tools it offered. Approve or block it with a reason; a blocked server is refused for all agents.

See. Score GenAI apps on data handling, compliance, security and terms, then record your own assessment. The starter scores are editorial judgements to re-assess, not measurements.

See. Charts of request volume with blocked traffic, top models and provider mix, plus tokens and cost, over a range you choose. Export the hourly series as CSV.

See. Drawn from real traffic: personal-data kinds sent to a model against those found and held back, and kinds returned in answers. Impact analysis shows what a change would touch.

See. A live map of agents, the models they call, the declared tools and data they touch and who they delegate to, with attacker paths to your data lit up.

See. Agents enter four ways: registered by hand, found in code, discovered in traffic or attested from CI or the cloud. Approval needs an owner and a second approver.

See. The models your gateway offers, and the AI SDKs, agent frameworks and MCP libraries your code actually imports. The code-discovered list is the strong part today.

See. A bill of materials for each repository that separates ordinary libraries from the AI inside it, exportable as CycloneDX or SPDX, with approved baselines and drift alerts.

Prove. One scan per repository across 8 source-control providers: secrets, vulnerable dependencies, insecure code and personal data. False positives teach the scanner; findings become tickets and build gates.

Prove. Paste code or point at a repository and get findings mapped to OWASP web and LLM categories, each with a suggested fix. A quick check; serious repository work goes to Repo Security.

Govern. Decide which code may be fetched, who approved it, and keep an audit trail of each request. Source is read in memory; findings are kept, not the source.

Prove. Safe checks on any public site; deeper active, API and LLM checks only on domains you have proven you own. A build gate blocks only on new problems.

Prove. Paste a chatbot, API or session URL and get the exact attack, the exact reply, the verdict and the framework reference for each break.

Control. Model files are inspected, never executed, and a poisoned artifact is stopped at the gate. Containment of machine-speed agents is opt-in, on limits you set.

Control. Reusable profiles follow the agent. Try a change on real traffic without enforcing it, roll it out to a few agents, enforce, and roll it back at any step.

Prove. Intelligence is matched to your SBOM, models, agents and vendors and reviewed by a person. Each threat ends in a recorded decision: owned work, a watch, a simulation or a no.

Control. Per-agent rules, a scanner for poisoned tool descriptions, tool pinning and a tester you can run before anything goes live. Argument checks are built in.

Prove. Each agent's configured controls are assessed against the OWASP Top 10 for Agentic Applications, showing which risks are covered and which are open. It is a configured-controls assessment, not an attack.

Control. Give each agent a zone with limits you set: call rate, session length, tokens, tools and model hosts. Each request is measured and each breach recorded; refusal is a rule you switch on.

Prove. See if a prompt is an injection or a jailbreak, the technique it uses and how it maps to OWASP and MITRE ATLAS, then what your gateway actually stopped in the last day.

Control. Code a model hands an agent is read for what it does before the agent gets it. It is read, never run.

Control. A fleet control room: posture, violations, spend, kill-switch state and an agent leaderboard in one view, with the agent named behind each refusal.

Control. Design workflows that ask agents, branch on their answers, wait for a human decision and contain or escalate. Each agent step runs through the gateway as that agent.

Govern. Where each agent stands, what changed this month and who changed it, from an append-only history of its lifecycle states.

Govern. The controls in force over your agents, how the fleet behaved against them, and a trail of who changed what. Enforced means a rule on the request path reads it.

Control. Record an agent, or your own privileged session, for a set time with full prompts and responses, then replay it with the decision each guardrail made on each request.

Control. Each agent gets a cryptographic identity, post-quantum hybrid on open standards, with signed requests that cannot be replayed, delegation that only narrows and a revocation list anyone can check.

Control. Enrol the agents allowed through the zero-trust gate, with expiry dates and one-click revocation. Roll it out in monitor mode first, then enforce when you are ready.

Control. One place to review the just-in-time access granted to each agent, and catch grants that ran past their approval or still belong to an agent you shut down.

Control. One agent, one tool, until a date, approved by someone else and revocable at any time. Access that expires replaces standing privilege.

See. Recognise each agent by its behaviour, see when that behaviour moves, and find out which known agent an anonymous caller really is. A baseline needs the agent's traffic to pass the gateway.

Control. An agent's declared purpose is checked against what its trust zone allows before a sensitive operation, and each live tool call is checked against the same scope.

Govern. A live heat map and your top risks, with a maturity read judged from how your register is actually run rather than from a questionnaire.

Govern. Entries arrive from compliance gaps, audits, scans, proven attacks and intelligence, de-duplicated and auto-closed. Each risk has an owner and a review date, and is never silently re-scored.

Govern. Enter low, likely and high estimates and get an annual loss range with its tail. The numbers come from your estimates, never from the product.

Govern. An inventory of AI vendors you actually depend on, discovered from your gateway and app catalogue, flagging those in use but not assessed. The register itself is a manual list.

Govern. Outbound sync of the risk register to your GRC, ticketing and third-party risk tools, signed and replayable. It sends out only; nothing is written back.

Prove. Attacks run in-path through your real controls and are judged blocked, detected, missed or refused by the model. See which control was meant to stop each probe and what you can close.

Prove. Exposures ranked by validated reachability and business impact, with drift alerts when a defence regresses. One status reads incomplete when a source could not be read.

Control. Incident playbooks built as real workflows, rehearsed safely before they are needed, with an honest count of which steps can actually run.

Prove. A twin of your agents with its campaigns, findings and health on one page. Attacks and guardrail tests run against the twin rather than production.

Prove. Run a campaign against named agents through your real gateway. The record says what was attempted and what was not, with narratives for the board, CISO, engineers and regulator.

Prove. Scan findings are chained into the routes an attacker could actually walk, with the evidence behind each step and a fix plan that you decide on.

Prove. A library of AI attack scenarios mapped to OWASP and MITRE ATLAS, each runnable against your own agents. A newly published technique can be turned into tests the same day.

Prove. Attack a copy of your agents, not production, then try the guardrail fix on the copy and promote it only if it holds.

Prove. Real faults are injected into AI traffic, with automatic rollback if the experiment hurts more than you planned. A failed experiment opens owned work with a due date.

Prove. Each week the newest backup is restored into a scratch database and timed, and the recovery time and data loss are recorded. A restore drill is a test, not a live failover.

Prove. Measured from real runs: techniques exercised, exercisable and untestable are kept apart, so an untested technique is never counted as covered.

Govern. Activate the frameworks that apply, such as the EU AI Act, NIST AI RMF and India DPDP, each mapped to or assessed against. Collect fresh evidence on demand.

Govern. Controls mapped to the EU AI Act, NIST AI RMF, ISO/IEC 42001 and NIS2 and assessed from live signals. Controls without a runtime signal are reported not assessable and left out of the score.

Govern. Each control carries a status someone decided and a health label showing whether anything proves it is working now, with the reasons one click away.

Govern. Counted from the resources discovery finds in your connected AWS account and the configuration audit run against them, with a daily history that starts when recording starts.

Govern. Graded A to D by how it was obtained, verified by a second person, kept byte-exact, sealed daily with a trusted timestamp, and withdrawn with a reason, never deleted.

Govern. Branded PDFs across 11 report types for the board and the auditor, with second-person sign-off, schedules and a flag when a report is out of date.

Govern. Watch the regulator pages that matter to you. New circulars land as candidates for a person to read and confirm, never as unverified compliance content.

Govern. Choose where you operate and what kind of entity you are, and the applicable regulators' reporting duties load as a rulebook with deadline clocks. It is starting content for your counsel to review.

Govern. Each control gets a verdict and a stated reason against live platform state. A control with no runtime signal is shown as not assessable instead of being silently passed.

Govern. Control-to-control mappings between common frameworks, each with a stated confidence and rationale. A supporting view that shows where one control answers clauses in another framework.

Govern. Framework targets, audit plans, evidence expiry, regulatory reporting stages, finding due dates and risk-acceptance expiries, all built from the records you actually hold.

Govern. Run each engagement through independence declarations, two-person review and management sign-off, ending in a sealed, timestamped archive anyone can verify. Preparer is never reviewer.

Govern. Generate an AI policy or upload yours and align it. Policies are versioned with a hash of the text, accepted per version by named people, with an auditor proof pack.

Govern. Each acceptance records who decided, why and until when, lapses back into open work on its end date, and goes to the board when it exceeds appetite.

Control. Runtime AI security incidents are deduplicated, ticketed and closed with a resolution note, and measured by trend and mean time to resolve. A runtime board, not a full case manager.

Control. Set which files employees may upload to the Assistant, with a size cap and allowed file types, and choose to block sensitive content before it reaches a model.

Govern. For each connection, see what ColossalX reads, does and never does. Test it, rotate or revoke it from the row, and set up signed outbound and verified inbound webhooks.

Govern. Search templates across SIEM, cloud security, identity, DevOps, ticketing and more, and connect from the card. Depth varies: three SIEMs send natively and the rest receive by signed webhook.

Govern. Connect a SAML provider by pasting its metadata, with just-in-time provisioning and Microsoft Entra directory sync. People sign in with the accounts they already have.

Govern. Map directory groups to access levels once, with custom roles down to the page, MFA required by role and an allowed email-domain list. Each screen tailors itself to the role.

Govern. Create a key with a lifetime and only the scopes the API checks. A key can never exceed its maker's permissions, the secret shows once, and a CI preset is ready.

Govern. Run ColossalX's own AI features on a managed model, your own model and key, or a self-hosted endpoint, per feature, with a live test. Each internal AI call is recorded.

See. A browser sensor, rolled out by browser policy, finds AI tools on laptops including local models that no network log can see. It is tested never to read a prompt.

Govern. Each record is labelled demonstration data or created by your own use, so evaluation numbers can be trusted. Counts are shown per table.