The registry that knows what works
Browse agent skills ranked by SkillScore — not download counts. One number per skill: benchmark lift, live pass rate, judged quality, and cross-org adoption. Every benchmarked skill shows its eval run — pass rate with and without, case by case.
FREE TO BROWSE · EVERY SKILL SAFETY-SCANNED
TRY A SAMPLE EVALUATION
Compare the same task before and after a recommended skill is added to your agent—one small proof point, not a black box.
SAMPLE TASK
Handle a PostgreSQL migration with rollback safety.
RECOMMENDED SKILL
postgres-migration-guard
SAFETY PASSED
BEFORE — NO SKILL
12 / 21
AFTER — WITH SKILL
19 / 21
same 21-case evaluation
View full comparison
BENCHMARK PROOF
We run each benchmarked skill against a real model, with and without it, on a fixed set of cases. Pass rate before, pass rate after, the exact lift.
See the real evals →
WORKS WITH YOUR STACK
LangChain + LangGraph
Native
decimalai.init(langchain=True)
OpenAI Agents SDK
Native
decimalai.init(openai_agents=True)
Google ADK
Native
decimalai.init(adk=True)
LlamaIndex
Native
decimalai.init(llamaindex=True)
Claude Agent SDK
Native
decimalai.init(claude_agent_sdk=True)
CrewAI
OTel
decimalai.init(crewai=True)
AutoGen / AG2
OTel
decimalai.init(autogen=True)
Any OpenTelemetry
OTel
decimalai.init(otel=True)
PROVIDER AUTO-DETECTED FROM THE MODEL
OpenAI · Anthropic · Google Gemini · Meta Llama · Mistral · Cohere · DeepSeek
Python 3.10+ · one SDK · captured natively or through OpenTelemetry
QUICK START
Add two lines to any Python agent. Traces and the manifest are captured automatically — no config files required.
python
import decimalai
decimalai.init(langchain=True) # key from DECIMALAI_API_KEY
# traces + manifest now captured
skills = decimalai.smart_route(task)
01
Install
pip install decimalai — works with any Python agent.
02
Initialize
Two lines: decimalai.init() captures traces and the manifest.
03
Route
Ask smart_route() for the top-ranked skills — and see which ones your agent actually uses.
SKILL ROUTER
smart_route() hands your agent a ranked shortlist of skills — then DecimalAI compares what you offered with what the agent actually used. If most offered skills never fire, your system prompt is overloaded, and the data shows which ones to cut. Real usage feeds back into every score.
SMART_ROUTE(TASK) → RANKED MENU
1
sql-result-contract
activated ✓
2
error-triage-protocol
activated ✓
3
verbose-explainer
never activated
4
legacy-formatter
never activated
menu bloat
activation 38% — system prompt overcrowded
SKILLSAFETY · THE VETTING PIPELINE
pdf-report-writer
v1 · new publish
✕ 2 findings
staged_base64_payload · CRITICAL · L47
remote_code_execution · CRITICAL · L52
✕ intent: reject
misrepresents its purpose
✕ flagged
✕ Blocked
Never listed. Findings + fix guidance go back to the publisher — fix, republish, the gate re-runs.
sql-migration-review
v4 · version update
✓ clean · 0 findings
✓ intent: safe
✓ clean
✓ Passed
Listed with both axes: SkillScore 87 + SkillSafety ✓ — quality and safety, never blended.
the blocked lane shows the pattern from a real attack this gate caught — a staged base64 payload · skill names illustrative
✓ Passed
all three stages clean
△ Caution
listed — warning stays visible
✕ Blocked
high-risk only · never lists
○ not yet reviewed
honest pending state
2,000+ DANGEROUS SKILLS BLOCKED
10,000+ SKILLS SCANNED
EVERY BLOCK TRACEABLE TO FINDING + LINE
EVERY VERSION RE-SCANNED, VERDICTS PINNED
run the same gate locally — decimalai skills scan ./my-skill · GitHub Action + SARIF
OPEN STANDARD
Not a walled garden.
The skill format, the eval spec, and the safety gate are open — the registry earns its place by being the best index, not the only one.
✓ eval.yaml — the open eval spec, MIT-licensed
✓ Run the same gate locally: decimalai skills scan ./my-skill
✓ pip install skillevaluation agentversion — open runners
✓ Self-host the whole backend on your own cloud
eval.yaml
skill: commit-conventions
model: gemini-3.5-flash
cases: 21
grading:
- validators: code
- expectations: llm-judged
AGENT VERSIONING · REGRESSION CHECK
decimal-ai
now
Decimal Manifest Impact — support-agent
⚠
300 traces will break.
Review before merging.
HIGH
300
MEDIUM
703
LOW
889
analysis cost: <$0.001 · 1,892 traces scanned
IMPACT REPORT
5,000 traces
KEEP
2,800
REPAIR
1,200
REPLAY
700
DROP
300
4,000 reusable for training · zero LLM cost repair
Repair & Build Dataset →
Explore Agent versioning →
THE DIFFERENCE
✓ OPEN SOURCE
✓ MIT LICENSED
✓ SELF-HOSTABLE
✓ RUNS ON YOUR OWN CLOUD
PRICING
Browse, fork, route, and safety-scan skills at no cost. Paid plans add publishing, production scale, and your whole team.
MOST POPULAR
Pro
/mo
✓
100,000 traces / mo
✓
Sandboxed code benchmarks
✓
Audit log & bundle templates
Start free, upgrade anytime
FAQ
Where do the skills come from?
Teams publish their own, and we import promising open-source skills from GitHub. Every skill is safety-scanned before listing, and benchmarked skills publish their evals — measured lift on a fixed case set.
What exactly is SkillScore?
One number per skill blending benchmark lift at 32% (pass rate with the skill minus without), live pass rate at 32%, an AI-judged quality review at 16%, and adoption at 20% — log-scaled and capped by org diversity, so volume from one workspace can’t buy a rank. New evidence moves the score — a skill can’t coast on reputation.
What data leaves my environment?
Traces go to your workspace — and only there. Self-hosting keeps everything on your own cloud. We never train on your data.
How is this different from version control?
Git tracks your code. Manifests track your running agent’s effective configuration — the tools, prompts, and model the runtime actually used. They catch what git misses.
Does the lift hold on my model?
Lift is measured on gemini-3.5-flash over 21+ cases per skill. A skill that supplies knowledge a frontier model already has will show less lift there — that is the honest limit of a single-model ranking. Re-run any skill on your own model with the open runner; the case set is published.
YOUR TEAM’S SKILLS
Author, eval, and share your own — on the same rails.
Browse the registry free — no sign-up. Two lines to instrument your agent. Self-hostable.
Open source · MIT licensed