koboi-agent 0.4.0

Universal configurable AI agent framework — production-grade, YAML-driven, open-source ready.

koboi-agent

Configurable AI agent framework. YAML-driven config, async Python 3.10+, multi-provider LLM (OpenAI, Anthropic, Cloudflare).

Features

• Multi-provider LLM: OpenAI, Anthropic, Cloudflare Workers AI
• YAML-driven config with ${ENV_VAR} interpolation
• Built-in tools: calculator, filesystem, shell, web, memory, search, git, subagent, task
• Hook lifecycle: 15 event types for logging, guardrails, telemetry
• RAG pipeline: chunking (fixed/sentence/paragraph/semantic), retrieval (keyword/semantic/hybrid), augmentation
• Guardrails: input/output validation, rate limiting, approval workflows, policy engine
• Multi-agent orchestration: keyword/LLM/hybrid routing, sequential/parallel execution
• Context management: truncation, smart truncation, key facts, sliding window
• Sandboxed execution: pluggable passthrough/restricted backends (per-session workdir, network/rlimit isolation)
• MCP client (stdio + HTTP) and server support
• HTTP/SSE server & jobs: koboi serve — interactive SSE chat (HITL) + autonomous background jobs; API keys, ownership, idempotency, durable resume
• Evaluation: BFCL, GAIA, SWE-bench, RAGAS, DeepEval scorers
• Terminal UI (Textual): chat, command palette, diff view, session management

Quickstart

Install

pip install -e ".[dev,tui]"

Set your API key

cp .env.example .env
# Edit .env and set OPENAI_API_KEY

Run the CLI

koboi chat configs/simple_chat.yaml

Run programmatically

import asyncio
from koboi import KoboiAgent

async def main():
    async with KoboiAgent.from_config("configs/simple_chat.yaml") as agent:
        result = await agent.run("What is 2 + 2?")
        print(result.content)

asyncio.run(main())

Serving (HTTP/SSE)

Run koboi as a stateless HTTP service: interactive SSE chat (with human-in-the-loop approvals) and autonomous background jobs (durable resume). Requires the [api] extra.

pip install -e ".[api]"
koboi keys create                               # mint an API key (Bearer auth)
koboi serve configs/server_deploy.yaml --host 0.0.0.0 --port 8080

Then:

# interactive SSE chat (stream tokens + HITL approvals)
curl -N -H "Authorization: Bearer $KEY" -H "Content-Type: application/json" \
  -d '{"message":"What is 2+2?"}' http://localhost:8080/v1/chat/stream

# autonomous job (202 + poll / SSE replay)
curl -X POST -H "Authorization: Bearer $KEY" -H "Content-Type: application/json" \
  -d '{"message":"Summarize the Q3 report"}' http://localhost:8080/v1/jobs

Two paths, same composition: koboi serve <config> (built-in) or create_app(config, extra_tools=..., extra_hooks=..., approval_handler=...) (customize by code). See configs/server_simple.yaml / configs/server_deploy.yaml, koboi/server/CLAUDE.md, and docs/rest-sse-requirements.md. Self-host deploy via the bundled Dockerfile + docker-compose.yml (Cloudflare Tunnel).

Configuration

Agents are configured via YAML. Key sections:

agent:
  name: "my-agent"
  system_prompt: "You are helpful."
  max_iterations: 10
  mode: "chat"  # chat | plan | act | auto | yolo

llm:
  provider: "openai"        # openai | anthropic | cloudflare
  model: "gpt-4o-mini"
  api_key: "${OPENAI_API_KEY}"
  base_url: "${OPENAI_BASE_URL:}"

tools:
  builtin: [calculator, web_search, memory_store, memory_recall]
  custom:
    - module: "my_tools"

context:
  strategy: "sliding_window"  # noop | truncation | smart_truncation | key_facts | sliding_window
  max_context_tokens: 8000

rag:
  enabled: true
  chunker: "paragraph"       # fixed | sentence | paragraph
  retriever: "keyword"       # keyword | semantic | hybrid
  top_k: 10
  documents:
    - path: "./data/sample/product_catalog.md"

guardrails:
  input:
    max_length: 10000
  rate_limit:
    max_calls_per_minute: 20

harness:
  doom_loop:
    consecutive_identical_threshold: 3
  telemetry: true
  carryover: true

See configs/ for full examples and .claude/skills/yaml-config.md for the complete schema.

Testing

pytest                        # all tests
pytest tests/test_config.py   # single file
pytest -k "hook"              # by keyword
pytest --cov=koboi            # with coverage

Examples

examples/ contains 32 numbered scripts covering every feature, plus server_built_in.py / server_customize.py for HTTP serving:

Range	Features
01-04	Basic chat and tool use
05-08	Context management and RAG
09-10	MCP client/server
11-14	Policy, hooks, skills, custom tools
15-16	Multi-agent orchestration
17	Anthropic provider
18-20	Harness (telemetry, doom loop, carryover)
21-24	Evaluation, production setup, SWE-bench, config-driven orchestration
25-28	Subagent delegation, task management, benchmarks, custom RAG
29-32	Skills (enhanced), eval-test, tool selection, sandbox + resume
server_*	koboi serve (built-in) and create_app() (customize)

Run any example:

python examples/01_simple_chat.py              # automatic mode
python examples/01_simple_chat.py -m interactive  # interactive mode

Architecture

For a detailed architecture overview (agent loop lifecycle, hook system, tool pipeline, extension points), see docs/architecture.md.

KoboiAgent (facade.py) is the single entry point. It assembles:

• AgentCore (loop.py) -- async agent loop
• RetryClient (client.py) -- LLM HTTP transport with retry
• ToolRegistry (tools/) -- tool registration and execution
• HookChain (hooks/) -- lifecycle event dispatch (15 events)
• ContextManager (context/) -- context window strategies
• AugmentationStrategy (rag/) -- RAG pipeline
• Guardrails (guardrails/) -- input/output validation
• PolicyEngine (harness/) -- rule-based tool filtering
• SkillRegistry (skills/) -- skill discovery
• ModeManager (modes.py) -- chat/plan/act/auto modes
• TrustDatabase (trust.py) -- graduated permissions
• Sandbox (sandbox/) -- passthrough/restricted execution backends (per-session workdir, network/rlimit isolation)
• StepJournal (journal.py) -- per-iteration step journal for crash/redeploy resume
• Server (server/) -- FastAPI HTTP/SSE serving (interactive chat + autonomous jobs)
• Orchestrator (orchestration/) -- multi-agent coordination
• SubAgentManager (subagent.py) -- parallel sub-agent delegation
• MCP clients (mcp/) -- external tool servers

All subsystems are configured from a single YAML file via Config (config.py).

License

MIT