Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure
TypeScript
import type { InternalToolDefinition, InternalToolResult } from '../workflow/types'
import { CAPABILITY_MAP, isCapabilityConfigured } from './capabilityMap'
/**
* Returns the configured capabilities for the LLM.
*
* Output (booleans + static hints, NEVER values):
* {
* "version": "v1",
* "configured": { "github": true, "brave_search": false, "smtp": true },
* "hints": { "github": "Available as $GITHUB_TOKEN ...", ... }
* }
*
* Security invariants:
* - No code path returns a secret value.
* - Hints are static strings from CAPABILITY_MAP — never derived from a
* value. Unit-tested by fuzz: response body is asserted to NOT contain
* any substring of any ConfigStore secret value, even when secrets are
* set, even when an attacker-supplied capability name is requested.
* - LLM provider key (OPENAI_API_KEY etc.) is intentionally absent from
* CAPABILITY_MAP — it's not an LLM-visible capability.
*
* The env-getter argument is supplied by the caller so this module stays
* decoupled from ConfigStore wiring; in production it routes ConfigStore
* → process.env fallback.
*/
export type EnvGetter = (key: string) => string | undefined
export interface GetCapabilitiesResponse {
version: 'v1'
configured: Record<string, boolean>
hints: Record<string, string>
}
export function buildGetCapabilitiesResponse(getEnv: EnvGetter): GetCapabilitiesResponse {
const configured: Record<string, boolean> = {}
const hints: Record<string, string> = {}
for (const cap of CAPABILITY_MAP) {
const ok = isCapabilityConfigured(cap, getEnv)
configured[cap.name] = ok
// Show the hint for every capability whether configured or not — the LLM
// benefits from knowing which env var name unlocks an integration even
// when not yet set, so it can correctly direct the operator.
hints[cap.name] = cap.hint
}
return { version: 'v1', configured, hints }
}
/**
* Build the InternalToolDefinition. Bound to the supplied env-getter so it
* can be registered alongside the existing `clerum__generate_*` tools in
* StepMcpRouter (workflow path) and NativeToolRegistry (chat path).
*/
export function createGetCapabilitiesTool(getEnv: EnvGetter): InternalToolDefinition {
return {
name: 'clerum__get_capabilities',
description:
'Discover which third-party integrations are configured for this Host. ' +
'Returns a JSON object with `configured` (presence-only booleans) and ' +
'`hints` (static usage strings — how to call each integration via shell ' +
'using $VAR expansion). Never returns secret values.',
parameters: {
type: 'object',
properties: {},
additionalProperties: false,
},
async execute(): Promise<InternalToolResult> {
try {
const payload = buildGetCapabilitiesResponse(getEnv)
return { success: true, content: JSON.stringify(payload) }
} catch (err) {
return { success: false, error: err instanceof Error ? err.message : String(err) }
}
},
}
}