loupfeed agents

Open-source platform for your team's internal agents — coding and product management — reachable from the tools you already work in.

loupfeed agents is a self-hostable agents platform: install it once, connect your tools, and invoke specialized agents from every entry point your team already uses —

• loupfeedcode — the coding agent. @-mention it on a GitHub issue or PR (or from Slack / Linear); it clones the repo into an isolated sandbox, implements, validates, and opens a draft PR. A separate reviewer graph reviews PRs with a diff-anchored findings model, and an analyzer graph learns your repo's review style over time.
• loupfeedpm (in development) — the product-management agent. Answers questions and acts on your planning system (Jira/Confluence via the Atlassian Rovo MCP connector; Notion, Linear, and others later), triages ideas out of meeting transcripts, and keeps planning threads moving — invoked from Microsoft Teams chats, channels, and call chats.

One platform underneath both: durable thread-scoped sessions (a conversation thread on the surface ⇄ an agent thread here), a single place to configure tools and connectors, BYO-LLM, and git as a first-class citizen.

Architecture

Built on LangGraph + Deep Agents. Each graph is an agent; all graphs share the platform's sessions, connectors, and configuration:

Graph	Purpose
agent	The coding agent — plans, edits, tests, and opens PRs from an isolated sandbox.
reviewer	Read-only PR reviews with a single evolving findings model.
analyzer	Learns per-repo review style from history and its own finding outcomes.
pm (planned)	The PM agent — planning-system connectors (MCP), meeting-transcript triage, thread-native conversations.

Key properties:

• Isolated execution. Every task runs in its own sandbox — Modal, Daytona, Runloop, LangSmith, or local mode on your own box — with full permissions inside the boundary and nothing outside it. Each thread's sandbox persists across follow-up messages, auto-recreates when unreachable, and tasks run in parallel without queuing.
• Deterministic thread routing. The same issue / PR / conversation always routes back to the same agent thread. Message the agent while it's working — mid-run messages are injected before its next step.
• Curated tools, not accumulated ones. A small toolset (execute, fetch_url, http_request, source-channel replies) plus the built-in Deep Agents file/search/subagent tools. Optional server-side observability tools (Datadog MCP, LangSmith) load only for authorized users' runs, with credentials that never enter the sandbox.
• BYO-LLM. Model selection is configuration, not code: any OpenAI-compatible endpoint (OPENAI_BASE_URL), Anthropic, per-user and team defaults, and model fallbacks.
• Context from where the work lives. A repo-root AGENTS.md is injected into the system prompt; the full issue / thread history rides along on invocation.
• Org config stays out of the platform. Deployments customize via an overlay — an env file plus an optional prompt-extension markdown appended to the agent's working-environment guidance — never by forking. See deploy/sandbox/README.md for the contract.
• Web dashboard (ui/) — GitHub login, per-user model/profile settings, team defaults, enabled repos, review-style management, and an agents chat UI.

Getting started

• INSTALLATION.md — local dev (backend + dashboard), GitHub App creation, triggers, production deployment.
• deploy/sandbox/README.md — single-box AWS deployment (SANDBOX_TYPE=local): systemd unit, Secrets Manager-rendered env, GitHub App auth shims, and the org-overlay contract.
• CUSTOMIZATION.md — swap the sandbox, model, tools, triggers, system prompt, and middleware.

make install   # uv pip install -e .
make dev       # langgraph dev — serves all graphs + the FastAPI webhook app
make test      # pytest

Roadmap

The platform direction (multi-surface entry points, thread⇄session model, connector registry) is documented in docs/11-agents-platform.md of the loupfeed repo:

1. Microsoft Teams entry point — 1:1 chats, channel threads, and meeting chats; connector sign-in initiated from Teams.
2. pm graph — Atlassian Rovo MCP + git; generic connector registry (Notion, Linear, …).
3. Meeting intelligence — transcript-driven triage and classification after calls.
4. Unified install — one onboarding across GitHub + Teams with shared identity.

License & credits

MIT. Derived from Open SWE by LangChain — thanks to its authors for the foundation this platform builds on.