Three ways. Pick the one your stack already speaks.

1. Drop this into Claude Code

~/.claude/server.json (or Cursor → Settings → MCP Servers)

{
  "mcpServers": {
    "company-brain": {
      "url": "https://your-brain.com/api/mcp",
      "transport": "http"
    }
  }
}

2. Verify the agent can see your skills

Run any agent — it should list 'company-brain' as a resource server.

# Tool surface visible to the agent:
brain://skills/refund-decision-policy
brain://skills/db-connection-pool-exhausted
brain://skills/pricing-exception-approval
brain://skills/auth-jwks-rotation-spike
... 9 total skills

3. The agent calls the brain when relevant

No code changes. The agent picks the right skill based on the task automatically.

User: "How do we handle a $312 refund — package never arrived?"
Agent: → resources/list → finds refund-decision-policy
       → resources/read brain://skills/refund-decision-policy
       → walks the runbook
       → returns the decision (with citations)

List all skills

GET /api/skills

curl https://your-brain.com/api/skills | jq

# {
#   "ok": true,
#   "count": 9,
#   "skills": [
#     {
#       "slug": "refund-decision-policy",
#       "domain": "cx",
#       "version": "1.2.0",
#       "confidence": 0.91,
#       "url": "/api/skills/refund-decision-policy"
#     }, ...
#   ]
# }

Match a task to the right skill

POST /api/skills/match — the brain.match() primitive

curl -X POST https://your-brain.com/api/skills/match \
  -H "content-type: application/json" \
  -d '{ "task": "customer wants refund, package never arrived" }'

# {
#   "matched": {
#     "slug": "refund-decision-policy",
#     "confidence": 0.94
#   },
#   "candidates": [...]
# }

Fetch the full skill

GET /api/skills/:slug?format=raw

curl https://your-brain.com/api/skills/refund-decision-policy?format=raw

# Returns the raw SKILL.md contents — agent loads directly into its context.

1. Clone the brain into your agent's workspace

cd ~/.claude/skills
git submodule add https://github.com/your-org/your-company-brain
git submodule update --init

2. Or use the Anthropic skill installer

Compatible with the open SKILL.md spec. One file per skill.

gh skill install your-org/your-company-brain
# → resolves to ~/.claude/skills/your-company-brain/*.skill.md

3. Brain updates land via PR

Every skill change is a git commit. Lint blocks risky merges. Approval workflow per skill.

# Nightly: brain ingests new Slack threads
# Generates PR: 'Update refund-decision-policy.skill.md (cap raised to $500)'
# Coherence engine runs on PR: detects contradiction with HR doc
# Routes to @rachel for sign-off
# On merge: agents pulling via submodule auto-update on next pull