Internal
registry for your coding agent
extensions.

Observal is a self-hosted registry for your coding agent extensions with a built-in insight engine. Set up Observal, define the scope, and share your Skills, MCPs, and Agents with your peers.

$pipx install observal-clicopy
live from your registry
registry // livethis week
components published / week
15
+38% vs last month
Components published this week
15+38%
Skills in registry
128+12%
Active sessions
51+21%
Install success rate
94%+8%
Avg insight score
86+4%
installs/day
4k
components
51
install success
94%
§ 01 · agent registry

Discover, upload, and
install agents in seconds.

Browse the components in your registry, upload your own skills, MCPs, hooks, and prompts, and install complete agents across Claude Code, Cursor, Kiro, and more.
observal.local/registry

Agent Registry

12 agents · 38 components · 6 publishers
Trending
pr-reviewerv2.3

Automated code review. Leaves structured feedback on pull requests:...

@devteam
8478.4
CursorClaude Code
incident-analystv1.2

Root-cause analysis for production incidents. Reads logs, traces, and erro...

@sre
6127.9
Claude Code
test-writerv3.0

Generates comprehensive unit and integration tests with full coverage analysis.

@qa
12049.1
Claude CodeKiro
docs-generatorv1.0

Produces API documentation, README files, and inline docstrings from source.

@infodev
3896.2
CursorVS Code
§ 02 · install · zero config

One command.
You're in.

~/my-repo - observal
works across your stack
Claude Code
OTEL
MCPAgentsSkillsHooksSandboxRAG
Cursor
MCPAgentsSkillsHooksSandboxRAG
Kiro IDE
OTEL
MCPAgentsSkillsHooksSandbox
Gemini CLI
OTEL
MCPAgentsSkillsSandboxRAG
VS Code
MCPAgentsSandbox
Copilot CLI
OTEL
SkillsHooks
§ 03 · why observal

Build fast. Ship faster.
Stay secure.

Bundle skills, MCP servers, hooks, and prompts into versioned components, then let real usage show you which ones help.
01 adoption

Track what gets used

HIGHLIGHT

Every install and session is captured, so you can see which skills, MCPs, and prompts your team actually reaches for, and which ones can be retired.

14 days agotoday
0%
components active
+8 this week
02 cross-harness

One component, every tool

$ observal agent pull pr-reviewer ✓
Claude CodeCursorKiroGemini CLICopilot CLIVS Code
03 private

Self-hosted by default

✓Docker Compose up. Your infra, your rules.
✓Components, sessions, and data stay local.
✓Open source, Apache-2.0. No lock-in.
§ 04 · components

Upload and track
every skill and tool.

Five component types, one local registry. Package them into agents, version every change, and see what each one is doing.
Skills

Reusable capabilities and patterns your agents can call on across tasks.

MCP servers

Tools and data sources exposed over the Model Context Protocol.

Hooks

Pre- and post-event gates: run lint, tests, and policy checks automatically.

Prompts

System prompts and instructions, versioned and shared across the team.

Sandboxes

Isolated runtimes that execute code and tests in a safe, reproducible box.

$observal registry skill submit --git-url github.com/acme/auth-helper✓ submitted
versionedreviewedtracked
Observability

See exactly what your agents do in production.

Every session from Claude Code, Kiro, Copilot CLI, and VS Code is captured as a structured trace. Token usage, tool calls, and outcomes. Searchable. Sortable. Always on.

No sampling.Every session, every tool call. Not 1-in-10.
Cross-platform.Claude Code, Kiro, Copilot CLI, VS Code, Gemini CLI.
Live updates.Active sessions surface in real time as they run.
Traces
All 18● Active 2
All platforms
All time
SessionUserPlatformTokensToolsDurationStarted
sonnet-4 · 12 promptsjsmithClaude Code12.4k / 3.1k2347m2m ago
haiku · 3 promptsaleeKiro1.2 cr-12mjust now
sonnet-4 · 8 promptsmkClaude Code8.1k / 2.2k142h 14m3h ago
sonnet-3-5 · 19 promptsrbCopilot CLI-311h 8myesterday
opus · 4 promptsjsmithClaude Code41.2k / 8.3k618m2d ago
Showing 5 of 18 sessions
Agent Insights

Not a score. A report that tells you what to fix.

Observal analyzes your agent's sessions and produces a specific action report: what's going wrong, which metric proves it, and the exact system prompt change that fixes it.

What to fix.Named issues with the metric that proves each one.
How to fix it.Exact system prompt changes, not vague recommendations.
Why it matters.Each suggestion traces back to a specific observed failure.
Insights · tester
tester
v1.0.0 · 1 session · Apr 20 – May 4, 2026
concerning
At a Glance
Working:Cache efficiency is strong at 76% -the agent reuses context well across tool calls.
Hindering:Error rate of 75% across tool calls. Most failures are command_failed on Bash -the agent runs commands against files it hasn't confirmed exist.
Quick win:Add a single line to the system prompt: verify target file exists with a Read before any Bash command.
Sessions
1
Apr 20 – May 4
Error rate
75%
6 of 8 tool calls
Cost / session
$0.22
claude-sonnet-4-5
Cache efficiency
76%
130k tokens saved
Suggestions · 3 actions
Verify target file exists before Bashhigh
Add to your agent system prompt: "Before running any Bash command, verify the target file exists with a Read call first."
Why: Addresses the 75% error rate -6 errors across 8 tool calls, all command_failed on Bash.
Add Bash error handling and retry logichigh
Add to your agent system prompt: "If a Bash command fails, catch the error, diagnose the cause, and retry with a corrected command before giving up."
Why: command_failed errors are the primary source of friction. The agent currently stops or hallucinates a fix instead of recovering.
Cap session duration at 400 secondsmedium
Add to your agent system prompt: "If you have been working for more than 6 minutes without completing the task, stop and summarize what you have done and what remains."
Why: Avg session is 397s -at the p90. Sessions that run long drive cost without proportional output.
§ 05 · agentic coding

The agentic coding
paradigm shift.

You no longer write every line. You direct agents in loops, steer, and ship. Watch one execute commands on this very section.
agent session - live
1

Direct, don't type

Set the goal in natural language. The agent scopes and writes the work.

2

Steer the loop

The agent proposes, you correct. One sentence redirects it mid-flight. No code to undo.

3

Ship on your terms

The output reflects your intent, not the agent's assumptions. Every trace proves it.

own your stack

Self-hosted.
Open source. Yours.

Observal runs entirely on your own infrastructure. One Docker Compose command brings up the registry, analytics, and databases. Your components, your sessions, and your data all stay local.

✓
Self-host
Docker Compose up. Postgres + ClickHouse + Redis. Your data stays yours.
✓
Local registry
Upload, version, and track skills, MCPs, hooks, and prompts in-house.
✓
Open source
Apache-2.0. Inspect everything, fork anything. No lock-in.
✓
Cross-harness
One agent definition renders config for every supported tool.