██████╗ ██████╗ ███████╗███████╗██████╗ ██╗   ██╗ █████╗ ██╗
██╔═══██╗██╔══██╗██╔════╝██╔════╝██╔══██╗██║   ██║██╔══██╗██║
██║   ██║██████╔╝███████╗█████╗  ██████╔╝██║   ██║███████║██║
██║   ██║██╔══██╗╚════██║██╔══╝  ██╔══██╗╚██╗ ██╔╝██╔══██║██║
╚██████╔╝██████╔╝███████║███████╗██║  ██║ ╚████╔╝ ██║  ██║███████╗
 ╚═════╝ ╚═════╝ ╚══════╝╚══════╝╚═╝  ╚═╝  ╚═══╝  ╚═╝  ╚═╝╚══════╝

Observal is the control plane and system of record for internal AI components

If you find Observal useful, please consider giving it a star. It helps others discover the project and keeps development going.

---

What is Observal and what does it solve?

Observal is the control plane and system of record for internal AI components. Every tech-forward organization today creates internal Skills, Agents, MCP servers and other AI components to boost productivity. Though the creation of these components has been prolific, the adoption and usage of such components is sparse. Developer/AI users today end up creating their own version of AI components without reusing existing packages.

The cause is largely due to two problems:

1. Lack of a discoverability layer

   Organizations store their AI components and agents in siloed github repositories with little to no documentation. Users are not able to locate similar components and this results in multiple developers creating the same/similar components again.

2. Missing feedback loop

   Any software where usage patterns are not understood and the principle of user-centric development is violated tends to fade out. Such is the problem with development of MCPs, Skills and Agents. Developers publish and maintain these components with little visibility into how they're actually used. Additionally, AI failures don't trigger static error codes: they hallucinate or provide subtly incorrect answers. This leaves users clueless about what went wrong compounding the feedback problem.

Observal solves this by providing a centralized discovery layer for AI components alongside useful insights into AI usage patterns. It turns silent failures into actionable feedback, ensuring internal AI tools are continuously optimized for the people using them.

Observal supports Claude Code, Cursor, Kiro, Pi, Copilot, Codex, OpenCode, and other tools.

Why teams use Observal

• Package components into reusable agents: Bundle Skills, MCP servers, hooks, prompts, and sandboxes into one versioned unit.
• Run a governed registry: Review submissions, approve internal agents, inspect version diffs, and give developers one trusted place to install from.
• Render across multiple Coding IDE/CLI: Generate the correct config for each supported harness instead of maintaining separate setup instructions for every harness.
• Let agents use each other: Every approved resource is discoverable over the open ARD standard, and a running agent can hand part of its task to another approved agent, whether it lives in the registry or is a remote A2A service.
• Learn what works: Use real adoption and session data to find which agents, tools, prompts, and workflows are helping teams.
• Replay sessions when needed: Use traces as evidence for debugging, review, audits, and deeper analysis.

---

Supported harnesses

harness
Claude Code
Kiro
Cursor
Pi
Copilot (CLI & VS Code Extension)
Codex
OpenCode
Antigravity CLI
Goose

One command to install any agent into any supported harness. The config files are generated per-harness automatically.

---

Quick Start

Observal has two parts: a server (API + web UI + databases) you self-host, and a CLI you install on each developer machine.

1. Deploy the server

One-line install (requires Docker Engine ≥ 24.0 with Compose v2):

curl -fsSL https://raw.githubusercontent.com/Observal/Observal/main/install-server.sh | bash

This downloads a Docker Compose package, generates operator-owned secret files with restricted container-group access, binds published ports to loopback by default, pulls container images from GHCR, and starts the stack. With a terminal it runs guided setup; without a terminal the same command applies safe defaults automatically.

Deployment docs are linked directly from this README:

• Setup guide: fastest path from zero to a working stack
• Self-hosting overview: deployment models and operator docs
• Production deployment: hardened production topology
• Databases: Postgres, ClickHouse, migrations, retention
• Upgrades: safe upgrade and rollback flow
• Release verification: verify checksums, provenance, and signed tags
• Backup and restore: backup plan before upgrades

From source (for contributors):

git clone https://github.com/Observal/Observal.git && cd Observal
cp .env.example .env
make up

2. Install the CLI

Standalone binary (no Python required):

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/Observal/Observal/main/install.sh | bash

# Windows (PowerShell)
irm https://raw.githubusercontent.com/Observal/Observal/main/install.ps1 | iex

Connecting to an existing server? See the Setup guide.

Python (3.11+):

uv tool install observal-cli
# or: pipx install observal-cli

3. Connect your harness

observal auth login
observal doctor --patch

This authenticates with your server, detects your harness, installs telemetry hooks, starts capturing sessions automatically, and prepares it for agent installs and registry commands.

Once logged in, run /observal inside your harness and it takes the wheel. Pull agents, submit components, browse the registry, run diagnostics:

/observal pull security-auditor
/observal scan
/observal doctor

Or just tell your agent what you want and it figures out the right commands.

---

How Observal works

Agents are portable context packages

An agent bundles 5 component types into a single installable package: MCP servers, skills, hooks, prompts, and sandboxes. You define the agent once, publish it to the registry, and Observal generates the right config files for whichever supported harness the user runs.

observal pull security-auditor --harness pi

The registry is the distribution layer

The registry is where agents live. Admins review submissions, version diffs keep changes auditable, and one command installs an agent into any supported harness.

Insights close the loop

Real usage data flows back as reports: what's helping, what's getting in the way, and where to improve. Session traces provide the underlying evidence for debugging and auditing.

---

Agent Registry

Browse, search, and install agents with harness compatibility badges:

Build agents visually with live config preview for every harness:

Components library: MCPs, Skills, Hooks, Prompts, Sandboxes:

---

Discovery and Delegation (ARD + A2A)

Installing an agent ahead of time only helps with the tasks you planned for. Observal also lets a running agent find what the organization already has, mid-task, and hand work to other agents.

ARD: one search across everything approved. Observal implements Agentic Resource Discovery (v0.91). Every approved agent, MCP server, skill, hook, prompt and sandbox is published as an ARD entry with a permanent urn:air: identifier, a public manifest at /.well-known/ard.json, and the spec's search API at /api/v1/ard/search. Results are ranked by relevance only; approval, visibility, harness support and whether the resource can be used right now are separate fields. The same visibility rules as the registry apply, so a team-private agent is only found by that team.

A2A: agents hand tasks to agents. Observal speaks the Agent2Agent protocol (v1.0, with v0.3 compatibility). Every delegation is an A2A Task, whoever runs it:

• Registry agents run headless in a harness that supports it (Claude Code, Kiro, Cursor, Codex, OpenCode, Copilot CLI, Antigravity, Pi) inside a throwaway copy of your repository. Their answer comes back as a result, and any file changes come back as a patch that is never applied for you.
• Remote A2A agents (a service another team runs, built with any framework) are registered by their Agent Card URL, reviewed like any submission, and then called directly with the card the reviewer approved. Observal never proxies the traffic or stores their credentials.

flowchart LR
    A["Agent in your harness"] -- "find_agents" --> R["Observal registry<br/>(ARD search)"]
    A -- "delegate" --> T{"A2A Task"}
    T -- "registry agent" --> H["Headless harness<br/>in a throwaway copy"]
    T -- "remote agent" --> S["A2A service<br/>(approved Agent Card)"]
    H -- "answer + patch" --> A
    S -- "artifacts" --> A

Loading

Every agent you pull gets an observal-agents MCP server with four tools, find_agents, delegate, get_task and cancel_task, so it can do this without being told how. Delegation is limited to approved agents, stops after two levels, lets a delegated agent start at most three tasks of its own, and refuses loops. Each delegation is recorded against the calling session, so traces show which agents a session relied on.

Discover shows registry agents that can take a task and remote A2A agents side by side:

Delegate from any terminal or script. Here a remote A2A agent answers directly; a registry agent also hands back its file changes as a patch for you to review:

observal delegate find "review this branch for auth bugs"
observal delegate run acme/security-reviewer "Review src/auth on this branch for token leaks"
git apply ~/.observal/delegations/<task-id>/changes.patch     # only if you agree with it

observal registry a2a submit https://agents.acme.com --visibility team --team platform
observal discover inspect urn:air:agents.acme.com:a2a:incident-triage     # shows the card's Digest
observal registry a2a review urn:air:agents.acme.com:a2a:incident-triage --approve --digest sha256:...

Design decisions: ADR 0001 (ARD), ADR 0002 (A2A delegation). API: Discovery endpoints.

---

Agent Insights

AI-powered insight reports analyze usage patterns across all sessions, what's working, what's hindering, and quick wins. Powered by LiteLLM, works with any provider (Anthropic, OpenAI, Bedrock, Gemini, Azure, Ollama).

See Insights LLM Setup for configuration.

---

Session Replay

Full session overview with token counts, models, tools, and turn-by-turn timeline:

Every turn captured: user prompt, tool calls, thinking block, assistant response:

Drill into any span to see exact tool inputs and outputs:

---

Review and Governance

Admin review queue with full prompt inspection and approve/reject:

Side-by-side version diffs before approving a new release:

Leaderboard tracks top agents and components by downloads:

---

Open-source features

Audit logs, SAML SSO, SCIM provisioning, and the executive dashboard are included in the Apache-2.0 distribution.

Audit log with parameterized search:

---

Documentation

Full docs at docs.observal.io.

Start here for deployment and operations:

Need	Link
Fast local or source setup	SETUP.md
Self-hosting overview	docs/self-hosting/README.md
Production deployment	docs/self-hosting/production-deploy.md
Single-node deployment	docs/self-hosting/single-node-deploy.md
Docker Compose setup	docs/self-hosting/docker-compose.md
Databases and migrations	docs/self-hosting/databases.md
Upgrades	docs/self-hosting/upgrades.md
Backup and restore	docs/self-hosting/backup-and-restore.md

---

Tech Stack

Layer	Technology
Frontend	Vite 8, React 19, TanStack Router, Tailwind CSS 4, shadcn/ui
Backend	Python 3.11+, FastAPI, Strawberry GraphQL
Databases	PostgreSQL 16 (registry), ClickHouse (telemetry)
Queue	Redis + arq
CLI	Python, Typer, Rich
Telemetry	Session hooks, local transcript reconciliation, push-based ingest
Deployment	Docker Compose (10 services), Kubernetes (Helm)

Contributing

See CONTRIBUTING.md. The short version:

1. Fork and clone
2. make hooks to install pre-commit hooks
3. Create a feature branch
4. Run make lint and make test
5. Open a PR

See AGENTS.md for internal codebase context.

Community

GitHub Discussions for questions and ideas. Discord for chat. Open Issues for confirmed bugs.

Reporting Issues

observal doctor support bundle

Produces a redacted diagnostic archive. Review before sharing: observal doctor support inspect observal-support-*.tar.gz

For live debugging, Observal uses loguru-based dev logging (internally called "optic"). Stream logs with:

observal logs

Logs are written to ~/.observal/logs/dev.log and include structured context for every request, background job, and telemetry event.

Security

Report vulnerabilities via GitHub Private Vulnerability Reporting or email contact@observal.io. Do not open a public issue. See SECURITY.md.

License

Observal is licensed under the Apache License 2.0. See LICENSE.