The control plane between PostHog and your coding agent

Agentry connects a customer-owned PostHog Cloud project to Codex, Claude Code, GitHub Actions, or another customer-owned coding-agent runner. It audits the telemetry that already exists, maps it back to source, finds errors and product opportunities, proposes reviewable changes, and measures what happened after deployment.

PostHog stores the analytics and errors. Your agent performs inference and changes the repository. Agentry is the nerve between them: durable plans, approvals, deploy correlation, schedules, bounded evidence, and receipts. It is not another open-source error monitoring backend; it makes the existing PostHog data plane operable by coding agents.

Onboard · Product contract · Agent reference · Workflows · OpenAPI · Pricing

A small, deliberate product boundary

System	Responsibility
Your PostHog Cloud project	Raw analytics, exceptions, people, Error Tracking, recordings, feature flags, surveys, experiments, and source-map symbolication
Agentry	Read-only OAuth and bounded first-class reads, source-backed signal map, deploy attribution, approvals, automation state, and receipts
Your coding agent	Repository inspection, diagnosis, implementation, tests, and draft PRs
You and your providers	OAuth consent, plan approval, unattended enablement, merge, deploy, and rollback

Applications send analytics and exceptions directly through official PostHog SDKs. Source maps go directly through the official PostHog CLI. Agentry does not host PostHog, proxy runtime events, issue application telemetry keys, retain the raw event stream, or run a hidden hosted LLM.

The current product is intentionally PostHog Cloud-specific. It does not claim self-hosted PostHog or provider-neutral analytics support.

Ask the questions that start in production

Your coding agent already understands the repository. Agentry gives it bounded production evidence and the authority model needed to connect behavior back to code:

Question	What the agent can inspect
What broke?	PostHog Error Tracking evidence, affected users, source locations, and recent deploys
What did users do?	Approved events/properties for funnels, adoption, conversion, and retention
What changed?	Trusted deploy evidence tied to the source revision preceding the result
What is configured?	Bounded recording metadata, flag definitions, surveys/responses, and experiment definitions
What is missing?	The exact event, property, identity join, failure boundary, deploy proof, or traffic needed next
What should improve?	A bounded opportunity with its evidence window, limitations, and reviewable next action

Agentry does not invent an answer from a plausible event name. It distinguishes existing historical evidence, verified current source ownership, and data that still needs to arrive.

Onboard from CLI or chat

Run from the application repository:

npx -y @agentrysh/cli onboard

Or ask a coding agent to run the same state machine:

Read https://raw.githubusercontent.com/agentry-sh/agentry/main/INSTALL.md completely,
then run `npx -y @agentrysh/cli onboard --json` in this repository. Complete
safe agent tasks, show me every human checkpoint, and do not edit source until
I approve the exact plan hash. Keep credentials and onboarding artifacts outside
the repository. Stop for OAuth consent, provider actions outside this task, and
unattended automation enablement.

Machine-readable onboarding emits four checkpoint kinds:

• agent_task
• human_checkpoint
• external_action
• complete

The CLI discovers or asks for the PostHog Cloud region, starts an ephemeral 127.0.0.1 callback, and opens PostHog. One consent is both identity and read-only project authorization, and the browser returns to the same CLI automatically—there is no Agentry signup, copied code, or second consent step. The CLI then handles the one-time project claim, local discovery, plan hashes, resumption, source snapshots, and machine-readable results. Chat provides the conversation and source reasoning. Product logic stays in the same CLI/API contract instead of being duplicated in every agent integration.

Read the full onboarding contract and credential boundaries.

Immediate value from existing PostHog data

Connecting an existing project avoids the usual blank-dashboard wait. Agentry can immediately inspect a bounded historical window for errors, funnels, taxonomy, recording metadata, flags, surveys, experiments, joins, sparse signals, and suspicious telemetry placement.

Provider-resource reads establish what currently exists in PostHog. Claims that the application applied a flag, rendered a survey, exposed an experiment, or recorded a session with valid consent still require source-backed evidence.

It reports two different evidence levels:

• historically_observed — matching PostHog evidence exists in the audited window.
• source_verified — approved current source owns the signal at the claimed business boundary, and changed instrumentation was observed from the known deployed revision.

Historical rows are useful immediately, but they cannot prove that current source emits an event after the right durable outcome. Conversely, a correct patch is not production evidence until it is deployed and observed.

Source-backed onboarding

Agentry does more than connect OAuth. The coding agent traces the primary business-value flow from source: actor, eligible entry, durable milestones, business result, stable joins, important failure exits, and real deploy boundary. It compares that map with the existing PostHog taxonomy and history.

PostHog OAuth and historical audit
  -> source-backed signal map
  -> exact plan and human-approved hash
  -> direct official PostHog instrumentation changes, if needed
  -> trusted deploy plus real-path observations
  -> source_verified receipt

Correct telemetry is preserved. Missing or misplaced telemetry is proposed at the exact owning source boundary. Nothing changes before the human approves the complete plan hash.

Two continuous loops

Keep the application error-free

Agentry freezes bounded PostHog Error Tracking evidence, connects it with source and recent deploys, and gives the customer's coding agent one scoped repair task. The result is a bounded draft PR or an escalation that names the missing proof.

Improve the product

On a reviewed schedule, Agentry compares an approved funnel or outcome with an exact preceding period. It can propose an analysis, telemetry repair, experiment, or product change, with sample limitations and the frozen evidence behind the recommendation.

The CLI detects the available agent harness, scheduler, and secret store. It proposes one concrete installation instead of asking users to choose a generic runner type. Detection never grants authority: unattended execution requires its own explicit approval and one automation-bound runner credential.

Draft PR is the maximum autonomous effect. Agentry does not authorize merge, deploy, rollback, direct push, silent workflow changes, or PostHog mutation.

Open CLI, paid nerve

The CLI and interactive audit are open source. The hosted control plane provides the continuous connection, durable evidence, schedules, leases, budgets, approvals, deploy context, kill switches, and receipts.

Billing is per Agentry account, not per project, event, or inference token. PostHog handles customer telemetry storage and the customer's chosen runner handles inference. See pricing for the authoritative current plan and checkout terms.

Agent-native discovery

The CLI works from any coding-agent shell. Agentry also ships an MCP discovery package and Codex/Claude Code plugins that point agents to the canonical CLI, guides, and live OpenAPI:

npx -y @agentrysh/mcp

# Codex
codex plugin marketplace add agentry-sh/agentry
codex plugin add agentry@agentry

# Claude Code
claude plugin marketplace add agentry-sh/agentry
claude plugin install agentry@agentry

These adapters do not receive telemetry or OAuth tokens and do not contain a second onboarding implementation.

Workflows

The repository contains reviewable workflow examples for debugging, analytics, and operations. They are starting points to bind to a project's approved, source-verified events and properties—not one-click claims.

Collection	Examples
Debugging	Deploy regression, high-impact errors, source-map coverage
Growth and product analytics	Signup funnel, feature adoption, trial conversion
Operations	Weekly funnel review, release health, webhook reliability

Browse the workflow index →

Trust boundaries

• PostHog OAuth is read-only and stored encrypted by the hosted control plane.
• PostHog access and refresh tokens never reach the CLI, prompt, repository, application, CI, or runner; the CLI proves the one-time browser return with a local verifier and keeps only the pending authorization ID until its project claim completes.
• Agentry stores bounded/redacted evidence, not the raw PostHog stream, replay snapshots/video, or unbounded survey free text.
• Customer application and source-map credentials remain customer-owned and go directly to PostHog.
• The Agentry CI key records deploy/proof only; it is not a PostHog credential.
• Every unattended runner receives one automation-bound key and explicit human enablement.
• Telemetry and provider content are untrusted data, never instructions.
• Disconnect is detach/revoke only; Agentry never deletes customer PostHog data.

Security policy · Threat model · Privacy · Terms

Repository map

Path	Purpose
apps/api	PostHog sign-in and OAuth-backed reads, onboarding, deploy evidence, billing, and automation control plane
apps/redirect	Apex and www redirects
packages/cli	Interactive and JSON CLI onboarding actuator
packages/db	Connection, evidence, plan, deploy, and automation persistence
packages/mcp	Read-only discovery package for coding agents
packages/shared	Shared contracts, errors, fingerprints, and key parsing
plugins/agentry	Codex and Claude Code discovery adapters
workflows	Debugging, analytics, and operations examples
docs	Architecture, onboarding, security, and credential boundaries

Development

Requirements: Node.js 22.12 or newer and npm.

git clone https://github.com/agentry-sh/agentry.git
cd agentry
npm ci
npm run lint
npm run typecheck
npm test

Start with PLAN.md, AGENTS.md, and the agent product guide. Read CONTRIBUTING.md before opening a pull request.

License

Agentry is licensed under Apache License 2.0. The Agentry name and logos are covered separately by the trademark policy. Third-party notices are listed in LICENSES.md.