agentic-concierge

agentic-concierge

A quality-first agent orchestration framework for local LLM inference.

• Router + Supervisor decomposes tasks into capabilities, recruits the right specialist packs, and runs them.
• Specialist packs are modular and composable: engineering, research, enterprise research — or add your own.
• Local-first: Ollama is the default and primary backend; any OpenAI-compatible server works via config.
• Extensible via MCP: connect GitHub, Confluence, Jira, filesystem, and other tool servers with a single config entry — no custom Python required.
• Observable: structured runlogs, persistent cross-run index, real-time SSE streaming, OpenTelemetry traces.

---

Key features

---

Installation

Quick install — Linux binary (recommended for end users)

curl -fsSL https://raw.githubusercontent.com/ausmarton/agentic-concierge/main/install.sh | sh

Downloads a static musl binary (~5 MB) to ~/.local/bin/concierge. Supports x86_64 and aarch64 Linux. No Python, pip, or package manager required.

On first run the launcher:

1. Detects or downloads Python 3.12 via uv
2. Creates a managed venv at ~/.local/share/agentic-concierge/venv/
3. Installs agentic-concierge from PyPI
4. Exec-replaces itself with the Python binary (correct PID, transparent signal forwarding)

Keep the launcher up to date:

concierge --self-update

Install to a custom directory (e.g. for system-wide install):

CONCIERGE_INSTALL_DIR=/usr/local/bin \
  curl -fsSL https://raw.githubusercontent.com/ausmarton/agentic-concierge/main/install.sh | sh

---

From PyPI (developers / non-Linux)

pip install agentic-concierge

Install optional extras:

pip install "agentic-concierge[otel]"   # OpenTelemetry tracing
pip install "agentic-concierge[mcp]"    # MCP tool server support

Docker (batteries-included: Ollama + agentic-concierge)

# Clone the repo for the config and docker-compose file
git clone https://github.com/ausmarton/agentic-concierge.git
cd agentic-concierge

# Start Ollama + agentic-concierge (pulls qwen2.5:7b on first run)
docker compose up -d

# Run a task
curl -X POST http://localhost:8080/run \
  -H "Content-Type: application/json" \
  -d '{"prompt": "Create a file hello.txt with content Hello World", "pack": "engineering"}'

The docker-compose.yml includes an Ollama service with a health check, an agentic-concierge service, and a one-shot model-pull service that exits after pulling qwen2.5:7b.

To use a different model, edit examples/ollama.json and re-mount it via CONCIERGE_CONFIG_PATH.

From source

git clone https://github.com/ausmarton/agentic-concierge.git
cd agentic-concierge
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

---

Quick start (local Ollama)

1. System dependencies

sudo dnf install -y python3 python3-devel gcc gcc-c++ make cmake git ripgrep jq

2. Install and start Ollama

# Install (pick one)
curl -fsSL https://ollama.com/install.sh | sh          # official script
# OR: sudo dnf install -y ollama                       # Fedora package

# Start (if not already running as a service)
ollama serve

Pull a model (agentic-concierge auto-pulls qwen2.5:7b if no chat model is found, but pre-pulling is faster):

ollama pull qwen2.5:7b     # fast model (default)
ollama pull qwen2.5:14b    # quality model (optional)

Any other model works — set CONCIERGE_CONFIG_PATH to point at a config with your preferred model name.

3. Install agentic-concierge

pip install agentic-concierge
# or from source:
# cd /path/to/agentic-concierge && pip install -e .

4. Run

# Quick smoke test — creates a file and lists the workspace
concierge run "Create a file hello.txt with content Hello World, then list the workspace." --pack engineering

Stream events as they happen with --stream (shows tool calls, LLM steps, results in real-time):

concierge run "Build a Flask /health endpoint with a test" --pack engineering --stream

You should see a run directory path and JSON with "action": "final". Check:

• .concierge/runs/<run_id>/workspace/hello.txt — artifact
• .concierge/runs/<run_id>/runlog.jsonl — structured event log (tool calls, LLM responses, etc.)

---

CLI reference

concierge run PROMPT [OPTIONS]

  Run a task using a specialist pack.

  Options:
    --pack TEXT              Specialist ID (e.g. engineering, research).
                             Omit to let the router pick based on capabilities.
    --model-key TEXT         Which model entry to use from config [default: quality]
    --network-allowed / --no-network-allowed
                             Allow web tools (web_search, fetch_url) [default: enabled]
    --stream / -s            Stream run events to the terminal as they happen.
    --verbose                Enable DEBUG logging

concierge serve [OPTIONS]

  Start the HTTP API server.

  Options:
    --host TEXT  [default: 127.0.0.1]
    --port INT   [default: 8787]

concierge logs list [OPTIONS]

  List past runs (most recent first).

  Options:
    --workspace PATH   [default: .concierge]
    --limit N          [default: 20]

concierge logs show RUN_ID [OPTIONS]

  Pretty-print runlog events for a run.

  Options:
    --workspace PATH
    --kinds TEXT   Comma-separated event kinds to filter
                   (e.g. tool_call,tool_result)

concierge logs search QUERY [OPTIONS]

  Search the cross-run index.
  Uses semantic similarity when embedding_model is configured;
  falls back to keyword/substring matching otherwise.

  Options:
    --workspace PATH
    --limit N          [default: 10]

---

HTTP API

Start the server:

concierge serve
# or: uvicorn agentic_concierge.interfaces.http_api:app --host 0.0.0.0 --port 8787

GET /health

{"ok": true}

POST /run — blocking run

curl -X POST http://127.0.0.1:8787/run \
  -H "Content-Type: application/json" \
  -d '{"prompt": "Create ok.txt with content OK", "pack": "engineering"}'

Request body:

{
  "prompt": "your task",
  "pack": "engineering",       // optional; omit to auto-route
  "model_key": "quality",      // optional; default "quality"
  "network_allowed": true      // optional; default true
}

Response: the finish_task payload merged with a _meta field containing run_id, specialist_ids, workspace, model, etc.

POST /run/stream — Server-Sent Events

Streams run events in real-time as they happen:

curl -N -X POST http://127.0.0.1:8787/run/stream \
  -H "Content-Type: application/json" \
  -d '{"prompt": "Create ok.txt with content OK", "pack": "engineering"}'

Each event is a data: <json>\n\n SSE line. Event kinds:

Rate limiting

When CONCIERGE_RATE_LIMIT is set to a positive integer, the API enforces a per-IP sliding-window rate limit (requests per minute). GET /health is always exempt. Excess requests receive 429 Too Many Requests with a Retry-After header:

export CONCIERGE_RATE_LIMIT=60   # 60 requests per minute per IP (default: no limit)
concierge serve

API key authentication

When CONCIERGE_API_KEY is set, every endpoint except GET /health requires an Authorization: Bearer <key> header:

export CONCIERGE_API_KEY="your-strong-secret"
concierge serve

# Include the header in every request:
curl -X POST http://127.0.0.1:8787/run \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer your-strong-secret" \
  -d '{"prompt": "hello"}'

Leave CONCIERGE_API_KEY unset (default) to disable authentication — suitable for local use. Uses constant-time comparison (hmac.compare_digest) to prevent timing attacks.

GET /runs/{run_id}/status

curl http://127.0.0.1:8787/runs/abc123.../status

{"status": "completed", "run_id": "abc123...", "specialist_ids": ["engineering"], "task_force_mode": "sequential"}

Status values: running, completed. Returns HTTP 404 if the run ID is not found.

---

Configuration

Set CONCIERGE_CONFIG_PATH to a JSON or YAML file to override the defaults.

export CONCIERGE_CONFIG_PATH=/path/to/your/config.json

The default config uses Ollama at localhost:11434 with qwen2.5:7b (fast) and qwen2.5:14b (quality). Copy examples/ollama.json as a starting point.

Key config fields

{
  "models": {
    "fast":    {"base_url": "http://localhost:11434/v1", "model": "qwen2.5:7b",  "temperature": 0.1, "max_tokens": 1200},
    "quality": {"base_url": "http://localhost:11434/v1", "model": "qwen2.5:14b", "temperature": 0.1, "max_tokens": 2400}
  },
  "specialists": {
    "engineering": {
      "description": "Plan → implement → test → review → iterate.",
      "keywords":    ["build", "implement", "code", "python"],
      "workflow":    "engineering",
      "capabilities": ["code_execution", "file_io", "software_testing"]
    }
  },

  "routing_model_key": "fast",         // model used for LLM-based routing
  "task_force_mode": "sequential",     // "sequential" (default) or "parallel"

  "local_llm_ensure_available": true,  // start Ollama if unreachable
  "local_llm_start_cmd": ["ollama", "serve"],
  "auto_pull_if_missing": true,        // pull qwen2.5:7b when no model exists
  "auto_pull_model": "qwen2.5:7b",

  "run_index": {
    "embedding_model": "nomic-embed-text"   // enables semantic search; omit for keyword-only
  },

  "cloud_fallback": {
    "model_key": "cloud_quality",           // must exist in "models"
    "policy": "no_tool_calls"               // trigger: "no_tool_calls" | "malformed_args" | "always"
  },

  "telemetry": {
    "enabled": true,
    "exporter": "otlp",
    "otlp_endpoint": "http://localhost:4317"
  }
}

Using a non-Ollama backend

Any OpenAI-compatible endpoint works. Set backend: "generic" for cloud/vLLM/LiteLLM servers (skips Ollama-specific 400 retry logic):

"models": {
  "quality": {
    "base_url": "https://api.openai.com/v1",
    "model": "gpt-4o",
    "api_key": "sk-...",
    "backend": "generic"
  }
}

Set local_llm_ensure_available: false when you manage the server yourself (CI, cloud deployments, etc.).

MCP tool servers

Attach any MCP server to a specialist pack — no Python code required:

"specialists": {
  "engineering": {
    "description": "Engineering with GitHub access.",
    "workflow": "engineering",
    "capabilities": ["code_execution", "file_io", "github_search"],
    "mcp_servers": [
      {
        "name": "github",
        "transport": "stdio",
        "command": "npx",
        "args": ["--yes", "--", "@modelcontextprotocol/server-github"],
        "env": {"GITHUB_TOKEN": "${GITHUB_TOKEN}"}
      }
    ]
  }
}

Tools are auto-discovered at startup and prefixed mcp__github__<tool>. See docs/MCP_INTEGRATIONS.md for GitHub, Confluence, Jira, and filesystem examples.

Parallel task forces

Run multiple specialists concurrently for independent sub-tasks:

"task_force_mode": "parallel"

In sequential mode (default) each pack receives the previous pack's output as context. In parallel mode all packs run concurrently via asyncio.gather and results are merged.

Podman container isolation

"specialists": {
  "engineering": {
    "container_image": "python:3.12-slim"
  }
}

All shell tool calls execute inside an isolated Podman container with the workspace mounted at /workspace. Requires Podman installed and the image available locally.

---

Specialist packs

Built-in packs

* Requires network_allowed: true (default).

Adding a custom pack

Option A — config-driven (no core change required):

# mypackage/packs.py
from agentic_concierge.infrastructure.specialists.base import BaseSpecialistPack
from agentic_concierge.infrastructure.specialists.tool_defs import make_tool_def, make_finish_tool_def

def build_my_pack(workspace_path: str, network_allowed: bool):
    tools = {
        "my_tool": lambda args: {"result": "..."},
    }
    tool_definitions = [
        make_tool_def("my_tool", "Does something useful.", {"type": "object", "properties": {...}, "required": [...]}),
        make_finish_tool_def(),
    ]
    return BaseSpecialistPack(
        specialist_id="my_specialist",
        system_prompt="You are a ...",
        tool_map=tools,
        tool_definitions=tool_definitions,
        workspace_path=workspace_path,
    )

"specialists": {
  "my_specialist": {
    "description": "My custom specialist.",
    "workflow":    "my_specialist",
    "builder":     "mypackage.packs:build_my_pack",
    "capabilities": ["my_capability"]
  }
}

Option B — built-in: add your pack factory to infrastructure/specialists/, register in _DEFAULT_BUILDERS in registry.py, and add an entry to DEFAULT_CONFIG. See docs/ARCHITECTURE.md §5 for the full extension guide.

---

Runlog

Every run produces .concierge/runs/<run_id>/runlog.jsonl. Each line:

{"ts": 1708800000.123, "kind": "tool_call", "step": "step_0", "payload": {"tool": "shell", "args": {"cmd": "ls"}}}

Inspect with:

concierge logs show <run_id>
concierge logs show <run_id> --kinds tool_call,tool_result

---

Testing

Fast CI (no LLM required, ~4 seconds, 1377+ tests):

pip install -e ".[dev]"
pytest tests/ -k "not real_llm and not real_mcp and not podman" -q

Full validation (requires Ollama + a pulled model):

python scripts/validate_full.py

Ensures the LLM is reachable (starts it if needed via config), then runs all tests including real-LLM E2E tests. Use ollama pull qwen2.5:7b or set CONCIERGE_CONFIG_PATH to a config with a model you have.

Single E2E check:

python scripts/verify_working_real.py

Runs one engineering task end-to-end and asserts that tool_call/tool_result events exist and workspace artifacts are created.

Test markers:

---

Development

See CONTRIBUTING.md for the full contributor guide.

# Install dev dependencies (includes mcp, pytest, pytest-asyncio)
pip install -e ".[dev]"

# Optional: OpenTelemetry
pip install -e ".[otel]"

# Run fast tests
pytest tests/ -k "not real_llm and not real_mcp and not podman" -q

# Lint
ruff check src/ tests/

---

Documentation

---

License

MIT