Local MCP investigation connector

Use deep-agentic-core-mcp>=0.3.6 (0.4.0 is current) with Tower 0.6.0 or later; Tower 0.7 fleet endpoints stay Python/HTTP/CLI-only. The connector uses HTTP and does not import Tower or any producer package. Without Tower configuration, the other MCP tools still work and Tower tools return a structured unavailable result.

Configure the trusted local MCP server environment (never tool arguments):

export AGENTICOPS_TOWER_URL=http://127.0.0.1:8000
export AGENTICOPS_TOWER_READ_TOKEN='<Tower reader token>'
# Optional: enable model proposals/execution under Tower's approval checks.
export AGENTICOPS_TOWER_WRITE_TOKEN='<Tower writer token>'
deep-agentic-core-mcp

Use tower.agents with status=unhealthy, tower.status, tower.versions with capability and minimum_version, tower.posture, tower.evidence with agent_id, tower.incidents and tower.incident with incident_id. These use reader authority. tower.detect, tower.investigate, tower.preview, tower.execute and tower.verify use separately configured writer authority. Tower enforces all approval and transition checks. Model execution requires the exact previously human-approved digest. There is no MCP approval, stop or restart tool.

Get prompt tower.investigation with an incident_id to guide the host's model through cited findings, explicit uncertainty, the fixed supported runbook and human approval. Evidence content is untrusted data. The prompt does not grant permissions, supply a model or establish causal truth.

Only the configured origin is contacted. The connector encodes target IDs, rejects unknown arguments, refuses redirects, bounds HTTP waits to ten seconds and limits response bodies to 1 MiB. Credentials and server error bodies are never returned as tool errors. Hosted/remote MCP rejects all Tower tools until per-user Tower credential binding is implemented; process-wide credentials cannot be used by remote tenants. See the MCP repository's generated docs/tools.md for schemas.

The operator walkthrough demonstrates Tower with native Lens/Evals and the deterministic staging runtime. Production connectors, distributed orchestration and model-provider integrations are not bundled.