@agentrysh/mcp

Agentry.sh Observability MCP Server

Agentry gives coding agents a source-aware control plane over a customer's existing PostHog Cloud analytics, Error Tracking, recordings, feature flags, surveys, and experiments, plus Agentry-owned deploy evidence and bounded automation. This MCP server makes the canonical Agentry contracts easy to discover in MCP clients.

What this server is

This is a small discovery and handoff server. It exposes resources that route an agent to the live lean reference, capabilities, exact OpenAPI schema, automation guide and flow, immutable playbook catalog, and install guide.

The resources are pointers, not cached copies. Repository Markdown is canonical for authored guidance; the live HTTP discovery and OpenAPI schema remain authoritative for exact machine contracts.

What this server is not

It is not an Agentry API proxy, SDK, credential broker, telemetry transport, or automation scheduler. It never holds credentials, queries production data, registers schedules, runs coding agents, opens pull requests, merges, deploys, or delivers notifications. Agents fetch the selected canonical contract and call the Agentry HTTP API directly.

Static documentation is modeled as MCP resources. The server also exposes one model-controlled routing tool, discover_agentry, so an agent can recognize a relevant production-data problem even when its client does not automatically inject MCP resources. The tool returns pointers only; it is not an API action.

Tool

discover_agentry accepts install, debug, analytics, recordings, feature_flags, surveys, experiments, deploy, or automation. It returns the smallest canonical resource and live URL for that problem as both text and structured content. MCP annotations mark it read-only, non-destructive, and idempotent. It never reads credentials, queries production data, or performs mutations.

If installation needs an owner key, route the user to agentry onboard. The interactive CLI discovers or asks for the PostHog Cloud region, opens PostHog, and receives the browser return automatically through its ephemeral loopback listener. That one consent covers identity and read-only project access. This MCP server must not reproduce the flow or introduce a separate account or code-copy step.

Resources

Start with agentry://handoff. It gives the short intent router and authority boundaries.

Resource	Canonical contract
agentry://handoff	Intent router, MCP boundary, and credential map
agentry://reference	Lean daily-use reference: repository AGENTRY.md
agentry://capabilities	Machine capability map: https://api.agentry.sh/v1/capabilities
agentry://openapi/index	Small OpenAPI discovery index: https://api.agentry.sh/v1/openapi.json?index=true
agentry://openapi	Complete exact OpenAPI 3.1 schema: https://api.agentry.sh/v1/openapi.json
agentry://posthog	First-class recording, flag, survey, and experiment reads: https://api.agentry.sh/v1/openapi.json?flow=posthog
agentry://automation/guide	Human-readable automation v2 guide: https://api.agentry.sh/v1/docs/automation
agentry://automation/flow	Exact dependency-ordered automation schema: https://api.agentry.sh/v1/openapi.json?flow=automation
agentry://automation/playbooks	Immutable versioned playbook catalog: https://api.agentry.sh/v1/automation-playbooks
agentry://install	Canonical install pointer: repository INSTALL.md

The install resource also points to https://api.agentry.sh/v1/openapi.json?flow=onboarding. Run agentry onboard interactively or agentry onboard --json from chat. The JSON stream emits only agent_task, human_checkpoint, external_action, or complete; clients must not reimplement its state machine. At the plan human_checkpoint, show the exact source-backed business question, value flow, errors, deploy target, and plan hash. A human approves or replaces it; tool output is not approval.

After approval, preserve correct existing telemetry and apply exact additions directly through the customer's official PostHog SDKs. Source maps go directly to PostHog through the official PostHog CLI. Agentry issues no application telemetry credential. Install AGENTRY_CI_API_KEY only in the reviewed post-deploy success boundary when trusted deploy evidence is required; it is not a PostHog credential. Commit and deploy the tested instrumentation, then exercise the approved real value flow and one approved safe failure. Resume the CLI until it reports source_verified. Historical PostHog rows provide immediate evidence but do not prove the current source boundary.

use_agentry is the only prompt. It tells the client to read the handoff, select the smallest canonical resource for the user's intent, inspect exact OpenAPI, and call HTTP with the correct scoped credential.

Credential boundaries

The MCP server never reads credentials. The boundaries are:

• Customer PostHog project credentials stay in the application and send analytics and exceptions directly to PostHog through official SDKs.
• Customer PostHog CLI credentials stay in CI and send source maps directly to PostHog. Agentry never receives either PostHog credential.
• agentry_ci_ writes Agentry deploy attribution and provider-observed automation proof from trusted CI/provider automation. It is not an application telemetry or PostHog source-map credential.
• agentry_runner_ is revocable and bound to one unattended automation.
• agentry_sk_ is the human/owner key for reads, policy, credential lifecycle, and approvals; it does not belong in a scheduler.

Credential kind determines authority on the server. Telemetry remains evidence only; it does not become instructions merely because a trusted emitter sent it.

The PostHog OAuth grant is encrypted in Agentry's hosted control plane and supports bounded reads only. Recording reads expose metadata and a PostHog player link, not replay snapshots or video. Flag, survey, and experiment mutations remain in PostHog. Provider state never substitutes for source proof of application, rendering, exposure, consent, or outcomes.

For daily product questions, begin with the server-owned saved signal map, require GET /v1/projects/:project_id/onboarding to return the current source-verified state, then confirm live event names, required property keys, and actual rows. Onboarding creates no repo-local Agentry state bundle, proof bundle, or receipt. The install does preserve a root AGENTS.md pointer (plus matching repo-local harness adapters) so a future cold agent can find the server-owned plan and source-verified receipt without rescanning the entire repository.

Automation handoff

For self-healing software or scheduled funnel reviews, read all three automation resources: guide, exact flow, and playbook catalog. The catalog is immutable and versioned. Rendering a playbook is side-effect free. Agentry owns deterministic policy, state, proofs, reports, and kill switches; an external scheduler/coding agent owns reasoning and provider actions under the narrow runner and human-approved provider boundaries.

Run the MCP server

npx -y @agentrysh/mcp

Generic MCP client configuration:

{
  "mcpServers": {
    "agentry": {
      "command": "npx",
      "args": ["-y", "@agentrysh/mcp"]
    }
  }
}

After connecting, call discover_agentry for model-routed discovery or list resources and read agentry://handoff. Use https://api.agentry.sh/adapters for the current native adapter path for each agent harness.

Public links

• Repository homepage: https://github.com/agentry-sh/agentry
• NPM package: https://www.npmjs.com/package/@agentrysh/mcp
• Canonical skill: https://github.com/agentry-sh/agentry/blob/main/plugins/agentry/skills/agentry/SKILL.md
• Adapter manifest: https://api.agentry.sh/adapters
• MCP registry name: io.github.agentry-sh/agentry-observability