Questions about EDAMAME, answered

EDAMAME is the AI agent trust layer for every machine your agents run on: workstations, CI/CD runners and agent servers, cloud or self-hosted.

Trust is proven two ways. Security posture answers whether the machine is hardened and compliant. Runtime behavior answers whether the AI agent on it is doing what it declared and staying clear of attack patterns. EDAMAME then binds access to code, secrets, production and critical company resources to that proof, so a stolen key or token alone is not enough.

AI agents such as Cursor, Claude Code, Claude Desktop, Codex, OpenClaw and Hermes act with the same access as the human who runs them, often more. They read code and secrets, install packages and run commands faster than anyone can review.

Their reasoning can also be corrupted: prompt injection, tool poisoning, goal drift, or a compromised package pulled during a task. When that happens, the agent's own log still reads clean. EDAMAME watches the machine from outside the agent, so what really happened is recorded, explained and, when needed, turned into an access decision.

• Human builders and developers who want a hardened workstation, plain-language guidance and portable proof of their security posture.
• Platform and DevOps teams who run CI/CD runners and agent servers and need them hardened and observed.
• Security and engineering leaders who need one view of every workstation, runner and agent host, and access policies that follow the evidence.
• Companies working with contractors, consultants and BYOD devices that will never be enrolled in an MDM.

None of the three. Sandboxes and harnesses confine what an agent may do; EDAMAME tells you when an agent is still running unconfined next to one. EDR protects the endpoint against malware; EDAMAME watches what the AI agent does across the whole host and compares it with what the agent declared. MDM administers company-owned devices; EDAMAME covers managed and unmanaged machines without remote control.

EDAMAME runs alongside these tools and does not replace them. It also checks that your existing protections, such as antivirus and EDR, are present and active.

Anything that runs inside the agent, or between the agent and its tools, sits exactly where an attacker who hijacked the agent already stands. EDAMAME observes at the operating system: which programs ran, which files were opened, which connections were made and which credentials were touched.

The agent cannot argue that record away. Dismissing a finding, switching detection off or pausing the observer is reserved to a human, and an agent that weakens its own sandbox or retries a command its own deny rules refused is itself reported.

No. EDAMAME is reporting-only on the device: it does not remove admin rights, force a VDI or lock the machine down. Humans see the same facts the security team sees, fix issues locally with guided or one-click remediation, and keep working. The aim is to secure AI agents while accelerating humans.

• EDAMAME Security: the desktop and mobile app. It understands, fixes and proves the security posture of a device, discovers the AI agents on it and verifies their runtime behavior.
• EDAMAME Posture: the CLI for CI/CD runners and any server running AI agents, cloud or self-hosted. It brings the same posture checks and runtime evidence to headless machines.
• EDAMAME Hub: the fleet control plane at hub.edamame.tech. It inventories workstations, runners and agent hosts, rolls up posture and runtime findings, drives conditional access, and holds your team plan.
• EDAMAME Portal: your personal account, the managed AI service for the assistant, and API keys for headless use.

Start with EDAMAME Security on your own machine — it is free. Add EDAMAME Posture on CI/CD runners and agent servers. When you want fleet visibility, policies and conditional access for a team, connect the devices to EDAMAME Hub. Every product produces the same evidence, so you can start small and grow without changing tools.

• Security posture and score, with continuous checks against operating-system best practices and guided or one-click remediation.
• Identity posture, including exposure of your accounts in known breaches.
• Network scan of your local network, with exposed services explained in plain language.
• Process-level network traffic analysis and file-integrity monitoring on desktop, with machine-learning anomaly detection.
• AI agent discovery, AI governance checks and runtime verification on desktop.
• Compliance views and signed, portable posture reports you can share.
• The AI Security Assistant, which explains findings and proposes fixes, with the model of your choice.

Detection and the assistant start when you turn on Protection. Findings appear in Detections; the assistant's own actions appear in AI History.

Posture checks run as soon as EDAMAME is installed. Runtime detection and the assistant start when you turn on Protection in EDAMAME Security. Every finding appears in Detections, whether or not a model is connected; the assistant's own actions appear in AI History while a model is connected.

EDAMAME Posture runs the same posture checks headless, then keeps watching the machine while it works: processes, files, network traffic and the AI agents running on it. On a CI/CD runner it hardens the runner before secrets or repositories are touched, enforces a network egress allowlist for the job, and can fail or cancel the pipeline when posture, traffic or behavior is out of policy. On an agent server, such as an OpenClaw or Hermes host, it adds runtime verification and attack-pattern findings.

It is distributed as a free CLI and as ready-made CI/CD integrations. See CI/CD security and the documentation.

• One inventory of workstations, CI/CD runners and agent hosts, with separate views for interactive machines and automation.
• Fleet scores, failed checks, events and escalations, with history.
• Policies built on minimum scores, specific checks, compliance tags or groups.
• Conditional access that applies those policies to your identity provider, source-control platform, VPN, firewall or Netskope.
• The AI Governance page for every AI agent across the fleet (Enterprise plan).
• Compliance evidence for your auditors, including the Vanta integration.

Learn more on the Endpoint Posture page.

With your GitHub or Google account, at hub.edamame.tech. Devices then join your workspace from the EDAMAME app or CLI with the user's email and a one-time code — or through a pre-configuration delivered by your MDM — without MDM enrollment.

The AI Security Assistant explains what EDAMAME found in plain language, ranks what to fix first and proposes the remediation. In Review mode you confirm each fix; in Auto mode it fixes the safe items and escalates the rest. Fixes can be undone wherever a rollback exists. The assistant runs while Protection is on.

You choose the model: EDAMAME's managed AI service, your own Claude or OpenAI key, or a local model through Ollama. Detection never depends on the assistant.

EDAMAME is open-core. The security checks (the threat models), the EDAMAME Posture CLI, the system helper, the foundation and networking libraries, and the CI/CD actions are open source under the Apache 2.0 license on GitHub. The detection core they build on, the EDAMAME Security app interface, EDAMAME Hub and EDAMAME Portal are commercial.

EDAMAME discovers and observes Cursor, Claude Code, Claude Desktop, OpenAI Codex CLI, OpenClaw and Hermes, wherever they run: locally on a laptop, on a remote host, or headless on CI/CD runners and self-hosted servers. EDAMAME runs on the same machine as the agent — for an agent in a VM, a container or on a remote server, run EDAMAME Posture there.

Each agent shows whether it is installed, observed, or has a paused observer, and a discovered agent that nobody is watching becomes a finding.

No. EDAMAME observes from outside the agent, at the operating system, so there is no plugin or SDK to install into Cursor, Claude Code, Codex or any other agent — and nothing for the agent to switch off. Once EDAMAME runs on the machine, discovered agents are observed automatically. On desktop, observing agent sessions uses the EDAMAME helper.

Intent divergence is the gap between what an AI agent said it would do and what the machine actually did. EDAMAME correlates three things: what the human authorized, derived from the human's own prompts so there is no policy file to write; what the agent declared in its plan; and what the operating system observed — processes, files, network connections and credentials.

When the observed behavior has no declared explanation, EDAMAME scores the divergence on an evidence trail. That is how it catches attacks that look legitimate inside any single view, such as an agent following instructions hidden in a poisoned tool.

Intent divergence runs while Protection is on and needs a connected model, which reads the human's request and the agent's plan. The behavioral detectors work without one.

• Credential harvest: a process opening several credential stores at once, or reading another process's memory.
• Token exfiltration and sensitive-data egress: a secret or sensitive file read, then sent to an undeclared destination.
• Cloud metadata access: a process reaching the cloud instance-metadata service to steal credentials.
• Supply-chain behavior: an install script writing outside its package, or a dependency reaching the network after a legitimate install.
• File tampering, and an agent weakening its own controls or retrying a command its own deny rules refused.
• Sandbox escape attempts.
• Tool poisoning: hidden instructions in MCP tool descriptions, or tool definitions that change after approval, plus the divergence they cause.

Before each release, EDAMAME must detect the replayed host behavior of real incidents, including the supply-chain attacks on axios, tj-actions, litellm and Shai-Hulud.

Yes, for most of them. EDAMAME detects the technique, not the indicator: reading another process's memory, opening a credential file and then reaching the network, writing a persistence hook outside a package tree. A new campaign almost always recombines techniques that are already covered.

In our September 2026 review of 12 publicly documented incidents, 11 used a technique EDAMAME already detects, so they would have been caught with no update. The twelfth, the XZ Utils backdoor, is caught at exploitation rather than while it sits dormant.

Most new attacks need no update at all. When an attack introduces a new indicator, such as a newly known-bad destination or a new class of sensitive file, EDAMAME ships it as a cryptographically signed threat model, verified before use (from version 2.0.3) — no new binary, no app update, no restart. Only a genuinely new sensing capability requires a software release.

No. Agent discovery, posture checks, traffic anomaly detection and the behavioral attack-pattern detectors are deterministic and run locally without any model.

A model adds three things: reading the agent's declared plan for intent divergence, adjudicating context to reduce noise, and explaining findings. It can never clear a corroborated critical finding. When no model is connected, the divergence verdict says so instead of guessing.

EDAMAME does not sit in the agent's execution path, by design: anything that blocks from inside or beneath the agent is reachable by the attacker who controls it. EDAMAME responds where the response cannot be forged. A finding degrades the device score; when your conditional-access policy includes that check, EDAMAME Hub withdraws the access the device no longer deserves — in your identity provider, repository, VPN, firewall or Netskope policy. On CI/CD, the pipeline can be stopped, and from a finding a human can also stop the offending process.

If you already confine agents with a sandbox or harness, EDAMAME tells you when an agent is still running unconfined next to it.

EDAMAME detects whether a recognized governance harness is installed — AgentField, nono, Rippletide or Anthropic Sandbox Runtime. Agents with no harness become a posture check, and when a harness is present but an agent still runs unconfined with a dangerous amplifier, such as root available without a password, EDAMAME reports it as the most severe AI posture finding.

The blast radius is how far a compromised agent could go on the machine before any network connection: whether it runs unconfined, whether root is available without a password, which categories of secrets have appeared in its sessions, and which high-reach programs (ssh, scp, docker, shells) it has already spawned. EDAMAME shows it per agent with the fix that shrinks it, such as removing passwordless sudo or running the agent in a sandbox.

Yes. For each agent, EDAMAME inventories the MCP servers and tools it can reach, their transport and whether they are authenticated, and flags an unprotected MCP server. On the Enterprise plan, the AI Governance page in EDAMAME Hub lists every external MCP service across the fleet with its authentication state, for devices that share AI details.

Yes. EDAMAME Security shows each agent's sessions with their token use and run economics, and a self-augmentation score that shows how effectively humans use their agents. A hash-chained flight recorder keeps a tamper-evident record of security-relevant agent activity on the device.

Three steps on the same evidence. Discover every AI agent on every machine, approved or not, with its MCP servers, tools, harness and blast radius. Audit each AI check in the frameworks your auditors read. Approve specific exceptions and allowlists for what you accept. The fleet-wide view is the AI Governance page in EDAMAME Hub, built from the AI details each device agrees to share. AI governance is part of the Enterprise plan; the Free and Business plans do not include it. See AI Governance.

• An agent running while its observer is paused.
• An agent with a large blast radius: unconfined, with root available without a password or secrets appearing in its sessions.
• An agent running with no governance harness around it.
• An agent running unconfined although a harness is installed — the most severe, because a bypassed control looks safe.
• An agent exposing an unprotected MCP server.
• Unreviewed divergence, attack-pattern findings and escalated actions.

They roll up into the normal device score and policies, next to classic checks such as disk encryption and patching.

An exception is approved for one specific condition — for example, one agent running ssh — never for a whole agent and never by switching off the check. Everything else the check covers stays flagged, a new condition fails it again, and history is never rewritten, so your audit trail stays truthful.

Exceptions quiet alerts and the governance view. Access decisions still follow the check itself. Exceptions are approved on the AI Governance page in EDAMAME Hub, on the Enterprise plan.

Yes, on the Enterprise plan. AI governance allowlists name the agents and MCP servers you approve. They start in audit-only mode; switch one to restrict what is allowed to exist and any agent or MCP server not on the list is reported as not permitted. Allowlists can be written ahead of a rollout, scoped to a group or a device, and given an expiry. They rely on the AI details each device shares.

Only the Enterprise plan. AI governance — the fleet-wide AI Governance page in EDAMAME Hub with agent adoption, harness coverage, external MCP services and the devices and allowlists matrix, plus allowlists and exceptions — is not included in the Free or Business plans, where the page shows an "Available on the Enterprise plan" card. On every plan, EDAMAME Security still runs the AI Agent Posture checks on each device, and they count in its score. See pricing.

You define a policy in EDAMAME Hub: a minimum score, specific checks such as AI Agent Posture, compliance tags or groups. EDAMAME Hub is the source of truth for whether each device meets it, and pushes that verdict to the enforcement point you already own. When a device falls out of policy, its access is withdrawn; when it is back in compliance, access returns automatically.

EDAMAME never sits in the network path: no proxy, no TLS interception. See Zero Trust.

• Netskope: device tag consumed by your Device Classification and Real-time Policy — details.
• GitHub Enterprise Cloud: IP allow-list of compliant devices on web, Git, SSH and API — details.
• GitLab: IP allow-list of compliant devices, runners included — details.
• Microsoft Entra ID: Conditional Access named location of compliant devices (requires Entra Conditional Access licensing) — details.
• Google Cloud Identity: Context-Aware Access level of compliant devices — details.
• Fortinet FortiGate: address group of compliant devices, by MAC address — details.
• Tailscale: compliant devices authorized on the tailnet; devices that stop reporting are deauthorized — details.
• NetBird: compliant devices placed in an access group — details.

Anything else with a REST or GraphQL API can be connected through a custom integration.

The device score degrades, EDAMAME Hub records the reason, and the device loses the access the policy protects — it is removed from the provider's allow-list, or its Netskope tag is removed. The user sees exactly why on their own machine and can fix it locally; once compliant again, access is restored without anyone touching the provider console. Escalations stay visible to the security team for review.

SSO and MFA prove who is asking. They cannot tell whether the machine behind a valid token is still trustworthy, or whether the AI agent on it is still doing what a human asked. Git tokens, SSH keys and API keys also work outside login flows.

Binding access to live posture and agent behavior means a stolen token is refused when it is used from a machine that has not proven its posture. With Netskope, Tailscale, NetBird and FortiGate the check is per device; with GitHub, GitLab, Entra ID and Google it is enforced through the network addresses of compliant devices.

Yes. That is the point of EDAMAME: managed or unmanaged, company-owned or personal, the same posture and runtime proof gates access, without MDM enrollment or remote control of the device.

Yes. EDAMAME Hub stays the source of truth, and the integration engine can talk REST or GraphQL to the access control you already run: identity provider, SSE, VPN, repository or firewall. Contact us to scope it.

The Netskope integration is live and publicly documented, built with Netskope as a technology partner. Netskope discovers and controls AI usage on the network; EDAMAME supplies what happens on the machine — endpoint posture, AI Agent Posture, intent divergence and attack-pattern findings.

EDAMAME applies a device tag in your Netskope tenant to devices that meet your EDAMAME policy and removes it when they stop meeting it. Your Netskope Device Classification matches the tag and your Real-time Policy decides access. EDAMAME labels, Netskope enforces. The device runs the Netskope Client and EDAMAME with its helper. See EDAMAME for Netskope.

• Tags from Netskope Cloud Exchange and other tools are preserved: EDAMAME reads the tags already on a device and changes only its own. If all five tag slots are taken, it reports an error instead of removing another tool's tag.
• Only the device whose compliance changed is updated — there is no tenant-wide sweep.
• With an allow-on-tag policy, a device that was never checked gets no access by default.
• There is no proxy replacement and no new enforcement plane.

EDAMAME streams device posture into Vanta through Vanta's public Device Monitoring API, the same way an MDM would — but with a lightweight agent instead of device-control profiles. Auditors get continuous SOC 2 and ISO 27001 evidence from unmanaged, contractor and BYOD workstations (Mac and Windows), with no MDM rollout. See Compliance.

They bind repository access — web, Git over HTTPS and SSH, and API — to the posture and runtime proof of the machine making the request, including CI/CD runners. EDAMAME Hub also offers a GitHub audit that reviews your organization's audit log through the EDAMAME GitHub App, with read-only access. See GitHub and GitLab.

Microsoft Entra ID and Google Cloud Identity are supported natively for conditional access. Other identity providers can be connected through a custom integration using their REST or GraphQL API.

Fortinet FortiGate, Tailscale and NetBird are supported natively: EDAMAME Hub decides which devices may reach protected destinations, and the firewall or mesh VPN enforces it, staying in the path as it is today. Netskope covers the SSE side.

Yes. EDAMAME covers Cursor, Anthropic Claude Code and Claude Desktop, OpenAI Codex, OpenClaw and Hermes without any plugin inside the agent. It also recognizes the governance harnesses that confine them — AgentField, nono, Rippletide and Anthropic Sandbox Runtime — and reports agents that still run unconfined next to one.

Classic posture checks map to CIS Benchmarks, SOC 2 and ISO 27001. Every AI check also carries its OWASP GenAI references (Agentic Top 10 and LLM Top 10), MITRE ATLAS, Agentic Trust Controls, ISO/IEC 42001 and ISO 27001. The Agentic Trust Controls scorecard grades 61 controls in 12 domains and marks each as enforced, partial or monitoring only. These are mappings and evidence, not a certification.

Yes, for Slack and Telegram. EDAMAME Security on desktop and EDAMAME Posture can send findings to a Slack or Telegram channel when an AI model is configured; messages include the device hostname and public IP. EDAMAME Hub notifies by email and phone with configurable thresholds, and keeps an in-app notification center.

Yes. EDAMAME Security on desktop can expose a local MCP server, off by default, so an AI agent you trust can read posture, findings and recommendations and request remediations. It listens on the device only and requires pairing credentials or a pre-shared key. EDAMAME Posture can host it on servers and runners too.

Technology, channel and service partners are welcome. If your product enforces access, collects compliance evidence or runs AI agents, EDAMAME can feed it host and agent truth. Contact us with a short description of your product and use case.

Yes. EDAMAME provides continuous endpoint evidence — encryption, patching, screen lock, firewall, protections — for workstations, contractor and BYOD devices and CI/CD runners, plus AI-agent evidence, without device administration. Evidence reaches your auditors through EDAMAME Hub, or through Vanta for Mac and Windows computers. EDAMAME prepares you for the audit; the certification itself remains your auditor's call.

It provides evidence toward them. Every AI check carries its ISO/IEC 42001 clause references, and EDAMAME records what each AI agent actually did on the machine — the runtime question auditors and the EU AI Act's logging and traceability expectations converge on. EDAMAME is evidence and readiness, not a certification.

It grades the runtime-observable subset of MITRE ATLAS: 39 parent techniques across 12 tactics, graded 16 strong, 20 partial and 3 indirect, each with its telemetry rationale and live findings. It is a detection-engineering map, not the complete ATLAS matrix, a certification or a MITRE endorsement.

A human can generate a signed, verifiable report of their device's security posture and share it with an employer, a client or an auditor — real proof instead of screenshots. The report carries the scores, the status of every check, the remediation history and the device's identifiers, so the recipient knows exactly what was verified. It is especially useful for contractors and consultants who work across several clients. The underlying certification method is patented.

EDAMAME Hub keeps history up to 7 days on the Free plan and up to 365 days on the Business plan. In the app, detections are kept 30 days and open attack-pattern findings stay until a human dismisses them. See pricing.

• No one can remotely control, wipe or run commands on the device through EDAMAME.
• When a device connects to EDAMAME Hub, a consent screen lists exactly which information will be shared.
• Remediation runs locally on the device, and fixes can be undone wherever a rollback exists.
• The app's privacy page lists every outbound connection — such as breach checks for the email addresses you monitor, location and DNS lookups, and remediation advice — and analytics and crash reporting can be switched off.
• The security checks come from a public, open-source threat model that anyone can review.

No. EDAMAME is reporting-only: no remote wipe, no remote control, no remote commands. Your organization sees the posture and findings shared with EDAMAME Hub and decides what that device may access.

On a company device managed by an MDM, the organization can pre-configure EDAMAME (from version 2.0.2) and lock a few settings: Hub enrollment, Protection, sharing of AI details and local network scan consent.

What the consent screen lists. A Hub report contains the device identifier and hostname, the operating system and EDAMAME versions, network identifiers (public IP, approximate location, MAC address, VPN or ZTNA peer IDs), domain and username, the security score by category, the status of every check and the remediation history.

AI details are shared only with consent, or when your organization requires them on a managed device. They contain agent names, harnesses, risk amplifiers, MCP server names, program and sensitive-file names, the public destinations of attack findings and the categories of secrets seen — never transcripts, prompts, file contents, command arguments, environment values or secret values.

If you never connect a device to EDAMAME Hub, its posture is not shared with any organization.

Two flows. Remediation advice: descriptions of flagged findings — device settings, network sessions or breached accounts — are sent to EDAMAME's AI service to generate the advice you see. The AI Security Assistant: when you connect a model, prompts carry the context needed to explain a finding, such as process names, paths, destinations, command lines and excerpts of the agent conversation of up to 3,000 characters each.

From version 2.0.3, prompts are minimized before sending: your home folder is written as ~, agent identifiers are pseudonymous, other user accounts are replaced and known secret formats are masked. With EDAMAME's managed AI service, assistant prompts are processed by Microsoft Azure OpenAI Service and your account and device identifiers are not forwarded; with your own Claude or OpenAI key they go directly to that provider; with a local Ollama model nothing leaves the device. Every flow, with its retention, is listed in the app before you connect.

EDAMAME reads the agent's session on the device to compare what the agent declared with what the machine did. You choose what is stored and shown: metadata only, redacted content (the default) or full content. Conversations stay on the device and are never sent to EDAMAME Hub. If you connect a model, excerpts of up to 3,000 characters can be included in its prompts; with AI details shared, the Hub receives only the categories of secrets seen.

EDAMAME Technologies is a French company and processes personal data under the GDPR: you can access, rectify, erase, restrict and port your data, and withdraw consent at any time. Questions go to privacy@edamame.tech. See the privacy policy.

EDAMAME ships no kernel driver of its own: it relies on the operating system's own observation interfaces and does not intercept — no proxy, no TLS interception, and it never drops a packet or silently blocks a process. Threat-model updates are cryptographically signed and verified before use, data sent to EDAMAME services is encrypted in transit, and the security checks themselves are open source and reviewable.

Yes, for the device itself. Scoring, network scan, traffic capture, anomaly detection, the deterministic detectors and the local MCP server work offline with the built-in threat models, and a local Ollama model keeps the assistant offline too. EDAMAME Hub, EDAMAME Portal and cloud AI models need connectivity.

• EDAMAME Security: Windows, macOS, Linux (Debian, Ubuntu, Raspberry Pi OS), iOS and Android, including ChromeOS through Google Play.
• EDAMAME Posture: Linux (Debian and Ubuntu families, Alpine, and plain binaries for amd64, arm64 and armhf), macOS and Windows, on workstations, CI/CD runners and servers.
• CI/CD: GitHub Actions on Linux, macOS and Windows runners, GitLab CI, and Jenkins.

See Download.

On iOS and Android, EDAMAME Security covers posture visibility: security posture and score, identity posture, network scan, compliance and posture sharing, and phones can join EDAMAME Hub. With a connected model, the assistant runs on phones too. Network flow capture, AI agent runtime verification and runtime detection run on desktop: Windows, macOS and Linux.

• macOS: the signed installer, the Mac App Store or Homebrew.
• Windows: the signed all-in-one installer, the Microsoft Store or Chocolatey. Traffic capture uses Npcap.
• Linux: the APT package.
• iOS and Android: the App Store and Google Play.
• Runners and servers: the EDAMAME Posture CLI, or the GitHub and GitLab actions.

Start from the download page, then connect the device to EDAMAME Hub if your organization uses it.

The helper is a small, open-source system service that performs the checks, remediations and rollbacks that need elevated rights, and powers traffic capture and agent observation. On macOS it is required. The signed all-in-one installers include it; with the Mac App Store, the Microsoft Store, Homebrew or Chocolatey it is installed separately. Without it on Windows, system checks, agent observation and session capture are unavailable. On Linux it is built into EDAMAME Posture. Source on GitHub.

No. EDAMAME is installed by the user, so it reaches the machines MDM does not: personal laptops, contractor devices and CI/CD runners. If you already run an MDM, you can use it to deliver and pre-configure EDAMAME on desktops (from version 2.0.2) — published guides cover Intune, Configuration Manager, Jamf, Kandji, Mosyle, Workspace ONE and Ansible. The MDM delivers and configures EDAMAME; EDAMAME itself never takes remote control of the device.

Add the EDAMAME Posture step at the start of the job. It checks and hardens the runner, keeps watching the job's processes and network traffic, and can fail the pipeline when posture or behavior is out of policy. Ready-made actions exist for GitHub Actions and GitLab CI, with examples for Jenkins. See CI/CD security.

EDAMAME is lightweight and observe-only. It ships no kernel driver of its own and does not intercept traffic, so it never sits in the path of your work or your agents.

Installation takes minutes per machine and conditional access is configured once in EDAMAME Hub. Teams typically run a short pilot on a few workstations and runners, then extend to the fleet. Read how Northbridge did it on the customers page.

• Free: €0 per month — full analytics and conditional access, AI Agent Posture checks on every device, up to 4 users, 1 device per user, 7 days of history, community support. AI governance is not included.
• Business: €150 per month — includes 10 users and grows with you, up to 5 devices per user, 365 days of history, CI/CD support, AI Agent Posture checks on every device, premium support, and a 7-day free trial. Save 20% with annual billing. AI governance is not included.
• Enterprise: custom pricing — everything in Business, plus enterprise integrations and AI governance: the AI Governance page in EDAMAME Hub, allowlists and exceptions.

See pricing.

With your own Claude or OpenAI key, or a local Ollama model, there is no EDAMAME charge for the assistant. EDAMAME's managed AI service has a free tier of 500,000 tokens, Pro at €5 per month for 2 million tokens and Max at €20 per month for 10 million, managed in EDAMAME Portal; the assistant pauses when the allowance is used.

Yes. The Free plan costs €0 per month for up to 4 users, with full analytics, conditional access and the AI Agent Posture checks on every device. It does not include AI governance, which is part of the Enterprise plan. Upgrade when you need more users, longer history or CI/CD coverage.

Yes. The Business plan comes with a 7-day free trial. Start it from the pricing page; it opens EDAMAME Hub, where you create your workspace and connect your first devices. AI governance is not part of the Business plan; it comes with Enterprise.

Yes. EDAMAME Security is free to download and use on your own devices to understand, fix and prove your security posture.

Yes. You can upgrade or downgrade as your needs change, and add users as your team grows.

Yes. You can cancel your subscription; the specific terms are in our terms of service.

When you need the fleet-wide AI Governance page, enterprise integrations, specific security or procurement requirements, or a deployment across many teams. Contact sales.

EDAMAME Technologies is a French cybersecurity company founded by Frank Lyonnet, PhD, a former INRIA researcher and serial cybersecurity entrepreneur. Meet the team on the about page.

Yes. EDAMAME's decentralized posture certification is protected by French patent FR3146742, granted in April 2026, with international applications pending. It certifies the posture of an unmanaged device without the verifier accessing or controlling the device.

Yes. Runtime verification for AI agents is grounded in research with the LISTIC laboratory (CNRS), published in open access on HAL. The paper also states its limits: the approach detects an attack as it unfolds rather than preventing the first action.

Yes. Netskope's Field CTO described EDAMAME as a platform that observes AI agents from outside the agents themselves and gates access when intent diverges from activity, in the article Agents deserve a place in every Zero Trust strategy.

• EDAMAME Security: documentation
• EDAMAME Hub: documentation
• EDAMAME Portal: documentation

Free plans get community support, including the EDAMAME community Slack reachable from the app; Business and Enterprise plans get premium support. Write to support@edamame.tech, use Feedback in the app (it can attach logs with your consent), open an issue on GitHub, or reach the team through the contact page. Privacy questions go to privacy@edamame.tech.

EDAMAME Security, EDAMAME Hub and EDAMAME Portal are available in English and French.

Yes. Install EDAMAME Security for free to see it on your own machine, or contact us for a walkthrough of EDAMAME Hub, AI governance and conditional access with your stack.