llms.txt

# Linchpin

> Open-source, self-hostable runtime for managed AI agents. Apache-2.0. Run agents on your own infrastructure, against any cloud model via OpenRouter or local model via Ollama.

Linchpin is a runtime layer for managed AI agents — closer to "LangGraph + your own FastAPI + your own docker-py + your own event log, pre-wired" than a library. Three services plus Postgres 16, runs on a single VM via `docker compose up`. It is an open-source alternative to closed managed-agent platforms like Devin, Manus, Cognition's tools, and Anthropic's Managed Agents.

Disambiguation: this Linchpin (linchpin.work, linchpinhq/linchpin) is unrelated to the older CentOS-PaaS-SIG/linchpin project (Ansible-based multi-cloud orchestration, distinct purpose and codebase).

## Key facts

- License: Apache-2.0
- Stack: FastAPI (Python), React, Postgres 16, Docker
- Model providers: OpenRouter (cloud, ~200 models) and Ollama (local) only
- Self-hostable on a single VM (~4 GB RAM minimum, 16 GB recommended)
- Repo: https://github.com/linchpinhq/linchpin
- Site: https://linchpin.work
- Latest release: v0.7.0 — Dreams (async memory curation, research preview)

## Capabilities (v0.7)

- **Sessions + sandbox**: per-session Docker container with eight built-in tools (`bash`, `read`, `write`, `edit`, `glob`, `grep`, `web_fetch`, `web_search`). Networking modes: `none`, `unrestricted`, `limited` (per-environment allowlist).
- **Files API**: content-addressable upload + mount as session resources. Deliverables watcher captures anything the agent writes under `/mnt/session/outputs/`.
- **Memory stores** (v0.3): persistent workspace-scoped filesystem state mounted at `/mnt/memory/<name>/`. Path-addressed, 100 KB cap, immutable versions with 30-day retention.
- **Skills** (v0.4): packaged expertise (SKILL.md + scripts) mounted at `/mnt/skills/<name>/`. Progressive disclosure — metadata always in system prompt, instructions on trigger, scripts on demand. Max 8 per agent. CLI: `linchpin skill build`.
- **Multimodal input** (v0.5): images (JPEG/PNG/GIF/WebP) and documents (PDF/text/markdown) in `user.message.content[]`. Sources: base64, URL, or Files-API reference.
- **MCP transport** (v0.5): both stdio (local subprocess) and URL (HTTP-streamable, for SaaS MCPs like GitHub Copilot, Linear). Auth via vault credentials.
- **Git repositories** (v0.5): clone any HTTPS git remote into a session at the requested `mount_path`. Per-environment clone cache, token rotation, generic (GitHub / GitLab / Bitbucket / Gitea).
- **Outcomes** (v0.6): declare a goal at session create — free-text definition + weighted rubric (weights sum to 1.0) + agent grader. `POST /v1/sessions/{id}/outcome_evaluations` records grader results; auto-terminates the session on success threshold.
- **Multi-agent threads** (v0.6): coordinator agents spawn worker sub-sessions via `POST /v1/sessions/{id}/threads`. Each thread runs its own agent in its own sandbox; every event mirrors up into the coordinator's event stream with a `thread_id` tag. Nesting cap (default 3 levels) prevents runaway coordinator-of-coordinator chains.
- **Dreams** (v0.7, research preview): async memory curation. `POST /v1/dreams` reads an input memory store + filtered past-session slice and a configured "dreamer" agent produces a new memory store with the cruft removed. Input store untouched. Opt in by setting `LINCHPIN_DREAMER_AGENT_ID`.
- **Webhooks** (v0.2): HMAC-signed at-least-once delivery for session lifecycle + credential events (including thread events as of v0.6).
- **Credential vaults**: Fernet-encrypted; referenced by id from agent + MCP configs.
- **Event streaming**: append-only per-session log, SSE replay + live tail.

## Documentation

- [Architecture](https://linchpin.work/architecture): system diagram and component breakdown
- [Docs index](https://linchpin.work/docs): full docs, links to repo
- [Quickstart](https://linchpin.work/#quickstart): hello-world in 60 seconds

## Comparisons

- [Comparisons index](https://linchpin.work/compare): all side-by-side pages
- [Linchpin vs OpenHands](https://linchpin.work/compare/openhands): runtime/platform vs agent+UI
- [Linchpin vs Devin](https://linchpin.work/compare/devin): open-source self-hostable vs closed hosted AI engineer
- [Linchpin vs Manus](https://linchpin.work/compare/manus): open-source self-hostable vs closed hosted general-purpose agent
- [Linchpin vs Anthropic Managed Agents](https://linchpin.work/compare/anthropic-managed-agents): open-source self-hostable vs closed hosted
- [Linchpin vs LangGraph](https://linchpin.work/compare/langgraph): runtime vs Python state-machine library
- [Linchpin vs CrewAI](https://linchpin.work/compare/crewai): runtime vs role-based multi-agent library
- [Linchpin vs AutoGen](https://linchpin.work/compare/autogen): runtime vs Microsoft Research multi-agent library
- [Linchpin vs Aider](https://linchpin.work/compare/aider): runtime vs CLI coding agent
- [Linchpin vs Dify](https://linchpin.work/compare/dify): code-first agent runtime vs visual LLMOps platform

## Use cases

- [Use cases index](https://linchpin.work/use-cases): all patterns
- [Self-hosted AI agents](https://linchpin.work/use-cases/self-hosted-agents): regulated industries, private repos, internal tooling
- [Local-LLM agents with Ollama](https://linchpin.work/use-cases/local-llm-agents): privacy-first agents running on local models
- [Self-hosted Devin alternative](https://linchpin.work/use-cases/devin-alternative): open-source equivalent of a hosted AI engineer

## Reference

- [FAQ](https://linchpin.work/faq): common questions about Linchpin
- [Glossary](https://linchpin.work/glossary): definitions of managed-agent runtime terms
- [Blog](https://linchpin.work/blog): notes on agent runtimes and self-hosting

## Optional

- [What is a managed agent runtime?](https://linchpin.work/blog/what-is-a-managed-agent-runtime): defining the runtime layer between agent libraries and hosted services
- [Self-hosting AI agents on a single VM](https://linchpin.work/blog/self-hosting-ai-agents-single-vm): step-by-step deployment guide