Agent loop
shippingThe core ReAct loop — prompt, reason, tool, result, next step. Streamed, interruptible, and fully logged.
Features
Every feature carries a status —
shipping is in the binary today,experimental works but rough,planned not yet shipped.The important parts
These are the pieces that define Glue: a terminal agent that edits code, runs commands, keeps local sessions, and can operate the web.
File editing
shippingTargeted patches, multi-file edits, new files. Every change lands in the transcript as a diff summary.
Command execution
shippingRun shell commands through your $SHELL. Output is captured, streamed back, and written to the session log.
Web tools
shippingFetch, search, extract, PDF OCR, browser automation. The CDP backend is experimental. Web →
Sessions
shippingAppend-only JSONL logs under ~/.glue/sessions/. Resumable across runs. Replay UI is planned. How it works →
Also ships with
Everything else you need to wire Glue into your setup.
Models & providers
shippingCurated catalog. provider/model IDs. adapter: openai for any compatible endpoint. Credentials stay out of project config. Catalog →
Host runtime
shippingCommands run in your shell on your machine — fastest path, uses the tools you already have installed.
Docker sandbox
shippingEphemeral containers for risky work. Workspace mounted in, container torn down with the session. Sandbox polish is experimental.
Cloud runtimes
shippingOffload commands to a remote sandbox. Daytona, Sprites, and Modal are available today via the same RuntimeFactory adapter pattern. Runtimes →
Subagents
shippingDelegate a sub-task to a separate agent with its own context and tool set. Useful for parallel search or independent investigations.
Skills
shippingDiscoverable, runnable skill definitions the agent can invoke as first-class tools.
MCP integration
shippingTalk to any MCP server as a tool source. stdio, HTTP+SSE, and WebSocket transports. Bearer and OAuth 2.1 auth, namespaced tools, live status in the TUI.