DAA

DAA — Debugging Autonomous Agent

Your app breaks at 3am. DAA investigates the root cause and opens a pull request — while you sleep.

[](https://www.python.org/) [](./LICENSE) [](https://fastapi.tiangolo.com/) [](https://python.langchain.com/) [](./Dockerfile)

---

How it works

Error fires in your app at 3am
          ↓
DAA SDK catches it → sends to DAA (one line of code)
          ↓
SHA-256 deduplication — same error again? Suppressed silently.
New pattern? Agent wakes up.
          ↓
AI Agent investigates across 4 dimensions:
   • Git commits  →  what changed recently?
   • App logs     →  what's the stack trace?
   • Traces       →  which request triggered it?
   • Source code  →  AST navigation to the exact line
          ↓
Opens a Pull Request with:
   • A code fix
   • Root-cause explanation
   • Postmortem summary
          ↓
You wake up, review the PR, and merge.

---

⚡ Quickstart — Serverless Mode (1 Command)

The easiest way to run DAA is as a stateless webhook receiver via the standalone Docker image. No databases or queues required. It receives an alert, talks to the LLM, pushes a branch, and opens a PR.

Requirements: Docker, a free Gemini API key, and a GitHub Personal Access Token.

docker run -p 8000:8080 \
  -e LLM_PROVIDER=google \
  -e GEMINI_API_KEY="your_api_key_here" \
  -e LLM_MODEL="gemini-2.5-flash" \
  -e DAA_DB_PROVIDER=none \
  -e DAA_GIT_TOKEN="github_pat_..." \
  -e GIT_HOST="https://github.com" \
  -e GIT_ORG="your-github-username" \
  rutvej1/daa-standalone:latest

Then trigger a test incident (simulating a webhook from Prometheus/Sentry):

curl -X POST http://localhost:8000/ingest/prometheus \
  -d '{"status": "firing", "alerts": [{"labels": {"alertname": "TestCrash"}}]}'
# → DAA immediately wakes up, clones the repo, queries the LLM, and opens a PR

Want the full persistent stack (Postgres + RabbitMQ + UI)? See [DEPLOYMENT.md](./DEPLOYMENT.md).

---

🔌 Two Ways to Integrate

DAA is flexible. You can plug it into your existing alerting stack, or instrument your code directly.

1. Existing Log Aggregators & Webhooks (Recommended for Serverless)

You do not need to use our SDK or change your app code. You can just point your existing alerting tools (Sentry, Datadog, Prometheus, CloudWatch) to DAA's webhook endpoints.

• Auth: Run DAA with DAA_AUTH_ENABLED=false and secure it behind your own API Gateway, AWS IAM, or Cloudflare Tunnel.
• Dedup: Rely on your existing aggregator to group the errors, and let DAA handle the autonomous fixing.

2. The DAA SDK (Recommended for Full-Stack)

If you don't have centralized logging, use the DAA SDK.

• Requires deploying DAA with DAA_AUTH_ENABLED=true (usually via Docker Compose).
• You register your app, get a DAA_TOKEN, and the SDK securely pushes exceptions to DAA.

# pip install daa-sdk
from daa_sdk import DAAClient

daa = DAAClient()  # reads DAA_TOKEN from env
daa.report_exception(exception, app_name="my-service")

---

What makes DAA different

| | Traditional alerting | DAA | |--|--|--| | When it fires | Every time the error happens | Once per unique error pattern | | What it tells you | "Error occurred" | Root cause + code fix | | What you do | Investigate manually | Review a PR | | 3am pages | Every time | Only for genuinely new problems |

---

Features

| | | |--|--| | 🔁 Zero alert fatigue | SHA-256 fingerprint dedup + sliding-window cooldowns | | 🧠 4-dimension investigation | Git history · Logs · Traces · AST code navigation | | 🔒 Agent safety | Hard 8-tool-call budget cap — no runaway LLM costs | | 🔀 Any LLM | Gemini · GPT-4o · Claude · Vertex · Ollama (air-gapped) | | 👤 Human-in-the-Loop | Approve AI fixes before the PR lands | | 🔧 Any git forge | GitHub · GitLab · Gitea · Bitbucket | | 🌐 MCP compatible | Use DAA as a tool inside Claude Desktop or Cursor |

---

Deployment

| Mode | Best for | How | |------|----------|-----| | Single Docker container | Try it out, small teams | docker run -p 8000:8080 --env-file .env daa:latest | | Docker Compose | Self-hosted, persistent | daa redeploy | | Serverless | Cloud Run / Fargate, zero-ops | DAA_DB_PROVIDER=none DAA_GIT_MODE=api |

Full guide: [DEPLOYMENT.md](./DEPLOYMENT.md)

---

Architecture

DAA/
├── app/
│   ├── backend-api/    ← FastAPI: ingest, dedup, incident tracking
│   ├── python-agent/   ← LangChain ReAct SRE agent (the brain)
│   ├── admin-panel/    ← React dashboard
│   └── daa-sdk/        ← Python SDK (Node/Go/Java/Ruby/.NET community)
├── daa                 ← CLI tool (daa init / register / test / logs)
└── docs/               ← Documentation

---

CLI

daa init              # Guided setup: LLM key, git token, deployment mode
daa register          # Register an app and get its DAA_TOKEN
daa policy            # Set escalation threshold (e.g. 3 errors in 60s)
daa test              # Fire a synthetic error and watch the pipeline
daa logs              # View recent incidents
daa status            # Health check all containers
daa redeploy          # Rebuild and restart everything
daa config set-model  # Switch LLM provider/model without restarting

---

Security

Self-hosted and private by design. See [SECURITY.md](./SECURITY.md) for the full hardening guide.

• Credentials via environment variables only — never mounted as files into containers
• CORS restricted to an explicit allowlist (CORS_ALLOW_ORIGINS)
• Webhook endpoints verify DAA_API_KEY + HMAC-SHA256 (Sentry)
• LLM agent tool-call budget is hard-capped (no unbounded loops)
• Report vulnerabilities: GitHub Security Advisories

---

FAQ

Q: Is DAA safe to use? A: Yes, DAA is self-hosted and private by design. It runs in your own environment (e.g., via Docker) and uses the API keys you provide.

Q: Is my code safe? A: Absolutely. DAA only analyzes the specific context around a triggered error (such as the stack trace, related Git commits, and the specific lines of code in the AST). It does not scan or upload your entire codebase.

Q: Is my code disclosed to the creators of DAA? A: No. DAA has no central server that receives your code or data. Any code context is sent directly from your self-hosted DAA instance to the LLM provider you configure (like Google Gemini, OpenAI, or a local Ollama instance).

Q: Can it run entirely offline/air-gapped? A: Yes, you can configure DAA to use local LLMs (like Ollama) so no data ever leaves your network.

---

Contributing

[CONTRIBUTING.md](./CONTRIBUTING.md) · [SECURITY.md](./SECURITY.md) · [LICENSE](./LICENSE)

PRs welcome. For large changes, open an issue first to align on the approach.

---

Built for engineers who are tired of being paged at 3am for the same error twice.

⭐ Star this repo if it's useful · Report a bug · Request a feature

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: rutvej
• Source: rutvej/DAA
• License: MIT
• Homepage: https://rutvej.github.io/DAA/

Install and usage instructions live in the source repository linked above.