anjor 0.6.0

Observability and reliability platform for agentic AI systems

Anjor

AI agents fail silently. A tool times out, a schema drifts, the context window fills up — and you find out from a user complaint, not a dashboard.

Anjor fixes that. It intercepts your agent's HTTP traffic at the protocol layer and gives you full visibility into every LLM call, tool use, and MCP server interaction — latency, token usage, context window growth, schema drift, prompt changes — without changing a single line of your agent code. Beyond passive logging, it surfaces actionable intelligence: failure pattern clustering, token optimization suggestions, and per-tool quality grades (A–F).

One-line install. No cloud. No account required.

---

Dashboard

The overview gives a bulls-eye view across the whole platform — tool health, LLM cost, MCP server status, top failure patterns, and drift alerts at a glance.

More screenshots

LLM Usage — token consumption and cost by model with daily trend and cache savings

MCP Servers — per-server call volume, success rates, and tool breakdown

Intelligence — failure clusters, token optimization opportunities, and quality scores

Tools — latency percentiles, drift detection, and per-tool drill-down

Traces — multi-agent span trees and cross-agent attribution

---

Install

pip install anjor

---

Quickstart

1. Start the collector and dashboard (one command, one port):

anjor start
# Anjor collector  http://localhost:7843/health
# Anjor dashboard  http://localhost:7843/ui/

2. Add one line to your agent:

import anjor
anjor.patch()   # that's it — httpx is now instrumented

import anthropic
client = anthropic.Anthropic()
# make calls as normal — they're captured automatically

Open http://localhost:7843/ui/ to see the dashboard.

3. Query the API directly:

curl http://localhost:7843/health
curl http://localhost:7843/tools
curl http://localhost:7843/mcp
curl http://localhost:7843/intelligence/failures
curl http://localhost:7843/intelligence/quality/tools

---

What it captures

Signal	Details
Tool calls	Name, status (success/failure), failure type, latency
MCP servers	Per-server call volume, success rate, latency — parsed from mcp__server__tool naming
Schema fingerprints	SHA-256 structural hash of tool input/output shape
Schema drift	Field-level diff against the baseline for each tool
LLM calls	Model, latency, finish reason — Anthropic, OpenAI, and Gemini
Token usage	Input + output + cache_read + cache_write tokens per call
Context window	Tokens used vs model limit, utilisation %, per-trace growth
Cache savings	Prompt cache hit rate and estimated cost savings
Context hogs	Per-tool average output size, % of context consumed
System prompt drift	SHA-256 per agent — alerts when prompt changes between calls
Failure patterns	Clustered failure analysis with descriptions and fix suggestions
Token optimization	Tools consuming >5% of context window, cost savings estimates
Quality scores	Per-tool reliability/schema-stability/latency grade (A–F)
Run quality	Per-trace context efficiency, failure recovery, diversity grade
Multi-agent spans	Parent/child span linking across agent boundaries
Trace graphs	DAG reconstruction, topological order, cycle detection
Cross-agent attribution	Token usage and failure rate broken down per agent

---

Supported providers

Provider	SDK	Intercepted endpoint
Anthropic	anthropic	api.anthropic.com/v1/messages
OpenAI	openai	api.openai.com/v1/chat/completions
Google Gemini	google-generativeai	generativelanguage.googleapis.com/.../generateContent

All three providers are auto-detected — no configuration required.

---

MCP tool analytics

MCP tools are automatically identified by their naming convention — no extra configuration needed. Any tool whose name follows mcp__<server>__<tool> is grouped by server in the MCP dashboard:

mcp__github__create_pull_request   →  server: github,     tool: create_pull_request
mcp__filesystem__read_file         →  server: filesystem, tool: read_file
mcp__brave_search__web_search      →  server: brave_search, tool: web_search

The /mcp endpoint returns per-server and per-tool aggregates and supports a ?days=N filter.

---

API endpoints

Method	Path	Description
POST	/events	Ingest a tool/LLM/span event
GET	/tools	All tools with summary stats
GET	/tools/{name}	Tool detail (latency percentiles, drift)
GET	/mcp	MCP server and tool aggregates (?days=N)
GET	/llm	LLM call summary by model (?days=N)
GET	/llm/usage/daily	Daily token usage by model (?days=N)
GET	/calls	Paginated raw event log
GET	/traces	Trace list (newest first)
GET	/traces/{id}/graph	DAG graph for a single trace
GET	/health	Uptime, queue depth, db path
GET	/intelligence/failures	Failure clusters sorted by rate
GET	/intelligence/optimization	Token hog tools + savings estimates
GET	/intelligence/quality/tools	Per-tool quality scores + grade
GET	/intelligence/quality/runs	Per-trace run quality scores + grade
GET	/intelligence/attribution	Per-agent token and failure attribution

---

Configuration

Via environment variables:

ANJOR_DB_PATH=./my_project.db python my_agent.py
ANJOR_BATCH_SIZE=1 ANJOR_BATCH_INTERVAL_MS=100 python my_agent.py
ANJOR_LOG_LEVEL=DEBUG python my_agent.py

Via .anjor.toml in your project root:

db_path = "my_project.db"
batch_size = 10
batch_interval_ms = 200
log_level = "DEBUG"

Via code:

import anjor
from anjor.core.config import AnjorConfig

anjor.patch(config=AnjorConfig(db_path="my_project.db", batch_size=1))

---

Limitations

• requests library not intercepted — all three provider SDKs use httpx by default
• Streaming responses are not parsed; only non-streaming calls are captured
• No cloud sync, authentication, or team features

---

Development

git clone https://github.com/anjor-labs/anjor.git
cd anjor
pip install -e ".[dev]"
pytest --cov=anjor --cov-fail-under=95 -q
ruff check anjor/ tests/
mypy anjor/
anjor start

See CONTRIBUTING.md for full guidelines.

---

Documentation

• Quickstart — see it in action
• Architecture — layer diagram and design decisions

---

Contributing & Contact

• Bug reports / feature requests — open an issue
• Questions / ideas — start a discussion

License

MIT © Anjor Labs