Coddy Agent

Session goal: /goal <objective> keeps the agent working until a second model, judging the tool results rather than the agent's claims, finds the objective met and a read-only verifier confirms it in the files; a watchdog cuts stalled or looping turns, and the goal shows in the web UI and the console (Session goal and supervisor).

Feature worktrees: worktree_open moves the active session into the worktree of a branch, creating one from the fresh default branch (or stacked on the current commit) when there is none, worktree_list shows where every worktree stands, and worktree_exit brings the session back to the main checkout, removing a worktree that holds nothing of its own (uncommitted work an explicit removal discards is kept as a snapshot); a new worktree gets the ignored files .worktreeinclude names, and a subagent can work in a worktree of its own; /worktree does the same for the operator, and the web UI shows the branch under the main project, with the way back on the plate (Git worktrees).

Hooks: your own commands at every lifecycle point, in Claude Code's hooks.json shape, able to deny, approve or rewrite a tool call (Hooks).

MCP servers over stdio, streamable HTTP and SSE, from the global and project mcp.json files or the editor, with a trust gate for what arrives with a checkout; connected all at once and, in the console, after the first frame (MCP servers).

Mentions: @ points at a file anywhere on disk, a line range, a folder, another session, a rule, a subagent or a web page, with the same fuzzy completion in the console and the browser; each is read once into the message that names it, so the provider's prompt cache holds (Mentions).

Documentation built in: this documentation ships inside the binary - a reader in the web UI, F1 in the console, coddy docs in the shell - searched offline, and the agent looks itself up with its own coddy_docs_search and coddy_docs_read tools or a @coddy:<page> mention (Built-in documentation).

Pictures in a prompt: paste a screenshot with ctrl+v in the console or into the web composer and mark it up first - pencil, ellipse, rectangle, arrow, text, crop, rotate and undo - in a window on the desktop or over the chat (Pasting and marking up pictures).

Message queue: steer a running turn at its next step or queue a separate turn after it, with images attached; every browser and console viewing the session sees the same queue (Message queue).

Background tasks: detached commands and subagent runs that wake the agent when they finish, in the Tasks panel of the web UI and /tasks in the console (Background tasks).

Canvas: pages the agent writes run as live frames on a board in the Files window of the web UI, at phone, tablet and desktop sizes at once; point at an element to ask for a change, and the agent reads the browser's console itself. Every board is an Obsidian .canvas file of the workspace, so it is in git and keeps its notes, groups and arrows (Canvas).

Preview server: "run this on a web server, I want to try it myself" - the agent serves the project directory on a free localhost port and hands you the link, no python -m http.server needed (Preview server).

Context compaction and long-term memory: /compact and automatic summarisation at a threshold, result eviction with keep_result, a memory subagent that runs per turn in the background, recalls what the notes hold and saves what you ask it to keep (Compaction, Memory).

Self-configuration: the agent edits its own YAML through staged config_* tools; nothing lands until you approve the commit (config.yaml reference).

Sessions everywhere: bundles on disk, resume from any surface, history rewind on message edit, /export to Markdown, HTML or JSON (Sessions, Session export).

Any model: OpenAI, Anthropic, NeuralDeep, ChatGPT through Codex, the models of a Devin account (Devin), a local ACP agent such as Claude Code, Codex or OpenCode served as a model (ACP agent as a model), Ollama, llama.cpp, vLLM and every OpenAI-compatible API, with reasoning levels and multimodal attachments per model and retries that honour Retry-After (Configuration).

Usage and limits: the quota behind the current model in the console footer and in the web UI for providers that publish one (NeuralDeep, Codex and Devin today), and a turn that hits a limit can wait for the reset and resume by itself (Console, config.yaml reference).

Config check, dry run and hot reload: coddy -t validates config.yaml against the embedded schema with file:line:col and a fix line, --dry-run probes paths, providers and models, ports, MCP servers, the Telegram token, remotes and swarm joins before anything starts, and a running coddy serve picks up an edited file without a restart (Configuration, coddy serve).

Fleets: a swarm relay lists and reaches many nodes, mounts their sessions into one list, routes across rings and pulls in nodes that can only dial out through a reverse tunnel; --remote drives a coddy serve on another machine from the console, an editor or the browser (Swarm, Remote mode).

In the browser: seven themes, English and Russian, @ mentions and multimodal attachments in the composer, a History grouped by folder with tags, pins and an archive, a plan document card, a Tasks panel, the documentation reader, a scheduler editor and a swarm map with the topology; it installs as an app and notifies you when the agent finishes or waits for you (Web UI).

Closed when it is on a network: a bearer token gates the API for clients, and an optional password sign-in gates the browser - coddy serve set-password, or CODDY_HTTP_USER / CODDY_HTTP_PASSWORD in ~/.coddy/.env - so a coddy serve on 0.0.0.0 is not readable by whoever finds the port (Remote mode, Security and trust).