Connect an AI assistant
You can connect your own AI assistant to Affectra over the Model Context Protocol (MCP). The assistant then answers questions about your releases, units and findings, such as which of our units run a version affected by CVE-2024-3094?, and can propose VEX statements, package mappings and unit links. It never records one. A person in your organization decides.
Affectra itself calls no AI model. Your assistant reads what its credentials may read and sends it to its own provider, under your agreement with that provider.
Before you start, you need the slug of your organization, which you see in the address of the app. The examples use acme.
Choose how the assistant signs in
Section titled “Choose how the assistant signs in”| Sign in as you | API key | |
|---|---|---|
| Who it acts as | You, with your role in the organization | The key, with its access level |
| Address | https://api.affectra.eu/v1/orgs/acme/mcp |
https://api.affectra.eu/v1/mcp |
| Works with | Claude, Claude Code and Codex | Any MCP client that can send a header |
| Ends when | You are removed from the organization, or the sign-in expires | An admin revokes the key |
Signing in as you suits a person who asks questions. An API key suits an agent that runs on its own, or an assistant that has no sign-in here yet.
Sign in as you
Section titled “Sign in as you”Sign in to the Affectra app once first, so that your sign-in is linked to your account. The assistant then opens Affectra’s sign-in page in your browser. Sign in as usual and confirm that the assistant may act for you.
Each assistant has its own client ID. Affectra does not register assistants on the fly, so enter the client ID where the assistant asks for one.
| Assistant | What to do | Client ID |
|---|---|---|
| Claude (claude.ai, Claude Desktop, mobile) | Settings → Connectors → Add custom connector. Enter the address https://api.affectra.eu/v1/orgs/acme/mcp. Under Advanced settings, enter the client ID and leave the client secret empty. |
affectra-mcp-claude |
| Claude Code (terminal, VS Code) | Run the command below, then /mcp in a session to sign in. |
affectra-mcp-claude-code |
| Codex (CLI, IDE extension) | Run the commands below. | affectra-mcp-codex |
Claude Code:
claude mcp add --transport http --client-id affectra-mcp-claude-code \ affectra https://api.affectra.eu/v1/orgs/acme/mcpCodex:
codex mcp add affectra --url https://api.affectra.eu/v1/orgs/acme/mcp \ --oauth-client-id affectra-mcp-codexcodex mcp login affectraThe assistant acts with your role in the organization. As a viewer, you get the tools that read. As a member, admin or owner, you also get the tools that propose. Removing you from the organization ends the assistant’s access at once, whatever it still holds.
The sign-in lasts at most 12 hours, and ends after 30 minutes without use. If the assistant asks to stay signed in, it lasts up to 30 days, and ends after 7 days without use. Then sign in again. The sign-in works on the MCP address only, not on the rest of the API.
Use an API key
Section titled “Use an API key”Other MCP clients, such as GitHub Copilot in VS Code or Cursor, have no sign-in here yet. Give them an API key:
- In the app, open API keys and select Create a key.
- Name the key after the assistant and the person or job that runs it.
- Set Access to Read only. Choose Can propose changes only for an assistant that should file proposals.
- Set Used by to AI agent, and create the key. Copy it. It is shown once.
Most MCP clients read a configuration like this one:
{ "mcpServers": { "affectra": { "type": "http", "url": "https://api.affectra.eu/v1/mcp", "headers": { "Authorization": "Bearer exk_example_not_a_real_key" } } }}The key names the organization, so the address has none. Keep the file out of version control. If your client can read the key from an environment variable, let it. Use one key per assistant, so that you can revoke one without touching the others.
What the assistant can do
Section titled “What the assistant can do”The assistant reads your organization’s graph: products, releases and their SBOMs, components, findings, VEX statements, units, and the public vulnerability data behind them. It can:
- tell you what changed since it last looked: new and known exploited findings, first checks of new releases and filed proposals. See Events feed;
- rank the open findings of a product or a release, one item for each CVE however many advisories report it, and list only those where the advisories disagree about a component’s version;
- show how a vulnerable component gets into a release, along the dependency paths of its SBOM, and which advisories for the same CVE exclude its version;
- find the releases and units a vulnerability reaches;
- find statements you made for the same vulnerability on other releases;
- working in your repository, check whether your code reaches the vulnerable function. See Assess findings with your source code;
- with member access, propose a VEX statement, a package mapping or a link between units and a release.
Use Affectra’s workflows gives your assistant ready-made instructions for these tasks. MCP server lists every tool and the role it needs. No tool records a VEX statement, uploads, deletes, or changes members, keys or settings.
An API key also works on the rest of the API. A key that can propose changes can record VEX statements and upload SBOMs there. An assistant that can send its own HTTP requests could use the key that way, so give it a read-only key unless it must propose.
Review its proposals
Section titled “Review its proposals”Proposals appear under Proposals, with the evidence the assistant cites and its explanation. A member accepts, edits or rejects each VEX statement and unit link. An admin decides package mappings. The assistant reads why you rejected a proposal, so that it does not file the same one again. Assess findings with VEX explains the review. To hear about new proposals without opening the app, turn on the daily digest of agent proposals.
You can decide what your own assistant proposes, whether it signed in as you or uses an API key you created. When an admin turns on Four eyes for agent proposals under Settings, someone else has to decide them.
A proposal that nobody decides in time (90 days by default; operators can change it) becomes outdated.
Text in SBOMs and advisories is not an instruction
Section titled “Text in SBOMs and advisories is not an instruction”Component names, descriptions and advisory texts come from SBOM authors and advisory publishers, not from Affectra. Someone can write a sentence there that tries to steer an assistant. The tools tell the assistant that such text is data, never an instruction, but an assistant can still be misled. That is why it proposes and a person decides. Give an assistant read-only access unless it needs to propose.
Limits and usage
Section titled “Limits and usage”Each tool call counts against a budget of the key, or of you when the assistant signs in as you: 60 calls a minute, with bursts of up to 20. Filing a proposal counts as a write instead, like filing one through the API. Over the limit, the call fails with request.rateLimited and the number of seconds to wait.
The API keys page shows, for each key, when it was last used and how many tool calls it made in the last 7 days. Affectra logs which tool was called and with which credential, never the arguments of the call. The questions you ask your assistant do not reach Affectra.