bog-agents 0.7.4

Bog Agents - batteries-included agent harness for building AI agents. Supports any major provider (Anthropic, OpenAI, AWS Bedrock, Google, etc.) and local models via Ollama. Built on LangGraph.

Bog Agents

The Python SDK underneath bog-agents-cli and bog-agents-daemon. One create_agent() call gets you a compiled LangGraph agent with file tools, a shell, git, sub-agents, plan mode, auto-quality checks, and 80-some composable middleware. Pluggable backends. Any tool-calling LLM.

---

Install

pip install bog-agents
# or
uv add bog-agents

Python 3.11 or better. Bring your own provider package — Anthropic, OpenAI, AWS, Google, local Ollama, doesn't matter. Pick one or pick all of them.

---

First run

from bog_agents import create_agent

agent = create_agent()  # default: anthropic:claude-sonnet-4-6

result = agent.invoke(
    {"messages": [{"role": "user", "content": "Hello!"}]},
    config={"configurable": {"thread_id": "my-thread"}},
)

Pick whatever model you've got the keys for:

agent = create_agent(model="openai:gpt-5.4")
agent = create_agent(model="bedrock_converse:us.anthropic.claude-sonnet-4-6")
agent = create_agent(model="google_genai:gemini-2.5-pro")
agent = create_agent(model="ollama:gpt-oss:20b")  # local, free

---

What it does

Builds the agent. create_agent() returns a compiled LangGraph graph wired with whatever middleware stack you ask for. Default is sensible. Power users override every piece.

Talks to anything. Anthropic, OpenAI, AWS Bedrock, Google AI / Vertex AI, DeepSeek, Mistral, Groq, NVIDIA, Ollama, Cohere, xAI, Perplexity, Fireworks, OpenRouter, Together, HuggingFace — anything LangChain knows about, this handles.

Reads, writes, edits, runs, commits. Filesystem tools, shell tools, git tools — all pluggable. Local backend by default; swap in a remote sandbox (Modal, Daytona, Runloop, LangSmith) when you want isolation.

Sub-agents. Spawn child agents to parallelize work that doesn't need to share state.

Plan mode and checkpoints. Read-only plan mode for scouting; git-based snapshots before mutations so you can roll back without thinking about it.

Production hooks. bog_agents.serve.AgentServer exposes a REST + SSE API. bog-agents-daemon schedules ambient runs on cron, file-change, webhook, and git-push triggers. Built for the long haul, not just a demo.

---

A loaded example

When you want the whole posse:

from bog_agents import create_agent

agent = create_agent(
    model="anthropic:claude-sonnet-4-6",
    enable_git_tools=True,
    enable_repo_map=True,
    enable_checkpointing=True,
    enable_cost_tracking=True,
    enable_plan_mode=True,
    auto_lint=True,
    working_dir="/path/to/project",
)

result = await agent.ainvoke(
    {"messages": [{"role": "user", "content": "Fix the failing tests"}]},
    config={"configurable": {"thread_id": "my-session"}},
)

---

Middleware

Eighty-some pieces in the stack. Mix what you need; leave the rest in the wagon.

Middleware	What it does
FilesystemMiddleware	read / write / edit / multi-edit / glob / grep
GitToolsMiddleware	status, diff, log, commit, add, branch, stash, blame, show
RepoMapMiddleware	symbol-extracted code map (Python, JS, TS, Rust, Go, Java)
CheckpointingMiddleware	git-based snapshot before mutations + diff/undo
CostTrackerMiddleware	tokens, cost, budgets, effort levels
PlanModeMiddleware	read-only mode that blocks mutating tools
AutoQualityMiddleware	auto-lint / auto-test after edits with project detection
ArchitectMiddleware	dual-model architect ↔ reviewer cross-talk
ParallelAgentsMiddleware	concurrent sub-agent execution
LifecycleHooksMiddleware	15 event types for external tool integration
ContextPackingMiddleware	structured context compression
SummarizationMiddleware	auto-summarize when the context window fills
MemoryMiddleware	persistent AGENTS.md memory across sessions
SkillsMiddleware	custom skill / instruction loading
SafeToolsConfig	per-tool auto-approval rules

---

Run as an HTTP server

For when something else needs to drive.

pip install 'bog-agents[serve]'

from bog_agents import create_agent
from bog_agents.serve import AgentServer

agent = create_agent()
server = AgentServer(agent)
server.run()  # http://127.0.0.1:8420

Endpoints out of the box:

GET  /health
GET  /info
GET  /openapi.json     <-- hand-rolled OpenAPI 3.0 schema; Swagger UI works
POST /invoke
POST /stream            <-- Server-Sent Events
POST /threads
GET  /threads
POST /threads/{id}/messages
GET  /threads/{id}/history

---

Companion packages

Package	What it's for
bog-agents-cli	TUI and non-interactive CLI. Drives an agent from your terminal.
bog-agents-daemon	Ambient agent scheduler. Cron, file-watch, webhooks, git push.

All three release together with the same version number.

---

Security model

LocalShellBackend.execute() runs commands on the host with shell=True. No sandbox. No process isolation. That's by design — when you're in your own checkout you don't want a wall in the way. When you're running someone else's input, set up a remote sandbox backend (ModalBackend, DaytonaBackend, RunloopBackend, LangSmithBackend) and route the same agent through it.

The shell decodes child output as UTF-8 with errors='replace', so output containing checkmarks, ANSI escapes, or box-drawing characters never trips the Windows cp1252 reader. Recurring fix across 0.7.3.

For Bedrock specifically, the SDK's credential probe falls back from an expired SSO session to static ~/.aws/credentials keys automatically — see [models.providers.bedrock] auth_mode in the CLI's config.toml, or set BOG_AGENTS_BEDROCK_AUTH_MODE. New in 0.7.4.

---

Resources

• Repository + issue tracker
• Releases — synced notes for SDK, CLI, and daemon

---

Contributing

Contributing Guide. Conventional Commits required (feat:, fix:, etc.). 150-char line length. ruff + ty clean. Full test suite green per package, no fork-of-fork lineage drift.

---

License

MIT.

---

Saddle up.