Aegis
Aegis

MCP setup

Aegis exposes a Model Context Protocol (MCP) server so AI agents — Claude Code, Cursor, Goose — can call Aegis tools directly. Nine tools cover health checks, job submission and inspection, error lookup, auth introspection, and load simulation.

Discover the config

The current SSE URL, tool list, and a drop-in client snippet live at a public endpoint:

GET https://radicalaegis.com/api/v1/mcp-config

You can also open Your account — there's an MCP card with a copy-paste-ready setup block for popular clients.

Account page MCP card
MCP setup card on the Account page — copy the snippet for your client.

Authentication

MCP uses your personal API key, passed as X-API-Key on the SSE connection. Create one on the API keys page. The same per-user rate limits apply.

One key, one session
Each SSE session is isolated by API key. Don't share keys between agents — revocation is per-key, and the access audit trail becomes meaningless when a key is shared.

Example: Claude Code

The config endpoint returns two fields you care about: config_snippet (a ready-to-paste MCP-server block) and setup_commands (the shell commands needed to wire it up). Fetch, inspect, then drop the snippet into your MCP servers file:

curl -s https://radicalaegis.com/api/v1/mcp-config \
  | jq '{config_snippet, setup_commands}'

What works via MCP

The MCP integration supports file submission and job monitoring. AI-generated summaries of high-risk results are not available through tool calls — this is an AI safety constraint that applies to all MCP clients, not specific to Aegis.

  • Submit a local file — ask your agent to call stage_fasta. It returns a bash command that uploads and submits the file directly; only the job ID and a results URL appear in the output.
  • Job status & history — list_jobs and get_job work for all results.
  • Health checks — check_health reports platform status.
Result interpretation
For AI-assisted summaries and interpretation, open the job results page and use the Ask Claude panel. The panel uses a different pathway that handles sensitive findings safely. MCP tool calls cannot relay this content.

Security model

See docs/MCP_SECURITY.md in the repo for the full threat model: per-user token store, LRU bounds, prompt-injection considerations for tool responses, and deployment guidance for operators.