anjor 0.3.0

Observability and reliability platform for agentic AI systems

Anjor

Observability and intelligence for AI agents. One-line install. No cloud. No account required.

Anjor intercepts your agent's HTTP traffic at the protocol layer and gives you full visibility into every LLM call and tool use — latency, token usage, context window growth, schema drift, prompt changes — without changing how you build. In Phase 3 it moves from passive logging to active recommendations: failure pattern clustering, token optimization suggestions, and per-tool quality scores.

---

Install

pip install anjor

---

Quickstart

1. Start the local collector (stores events to SQLite):

python scripts/start_collector.py

2. Add one line to your agent:

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

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

3. (Optional) Start the local dashboard:

bash scripts/start_dashboard.sh   # opens http://localhost:7844

4. Query data directly:

curl http://localhost:7843/health
curl http://localhost:7843/tools
curl http://localhost:7843/intelligence/failures    # Phase 3: failure patterns
curl http://localhost:7843/intelligence/quality/tools   # Phase 3: quality scores

No API key? Use respx to replay a mock response — see the quickstart guide.

---

What it captures

Signal	Phase	Details
Tool calls	1	Name, status (success/failure), failure type
Schema fingerprints	1	SHA-256 structural hash of tool input/output shape
Schema drift	1	Field-level diff against the baseline for each tool
Latency	1	Per-call and aggregated (p50/p95/p99)
LLM calls	2	Model, latency, finish reason — for every Anthropic /v1/messages call
Token usage	2	Input + output + cache_read tokens per call
Context window	2	Tokens used vs model limit, utilisation %, per-trace growth rate
Context hogs	2	Per-tool average output size, % of context consumed
System prompt drift	2	SHA-256 per agent — alerts when prompt changes between calls
Trace context	1–2	Trace ID, session ID, agent ID — consistent across LLM + tool events
Failure patterns	3	Clustered failure analysis with natural-language descriptions and fix suggestions
Token optimization	3	Tools consuming >5% of context window, estimated token waste and cost savings
Quality scores	3	Per-tool reliability/schema-stability/latency-consistency grade (A–F)
Run quality	3	Per-trace context efficiency, failure recovery, tool diversity grade (A–F)

---

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))

---

What is NOT in v0.3

• No multi-agent tracing (Phase 4)
• No cloud sync, authentication, or team management
• requests library not intercepted (Anthropic SDK uses httpx by default)
• OpenAI parser not implemented (stub only)
• Intelligence suggestions are heuristic — no LLM-powered explanations yet

---

Development

git clone https://github.com/anjor-labs/anjor.git
cd anjor
bash scripts/dev_setup.sh   # creates .venv, installs deps
source .venv/bin/activate
.venv/bin/pytest            # ≥95% coverage enforced
ruff check .                # zero lint errors
mypy anjor/            # strict type checking

See CONTRIBUTING.md for full guidelines.

---

Documentation

• Quickstart — see it in action
• Architecture — layer diagram and design decisions
• Code flow — execution traces end-to-end

---

License

MIT © Anjani Kumar