Repository navigation
Releases
AgentSquid 0.1.6
AgentSquid 0.1.6
Highlights since 0.1.5, including all 0.1.6 release candidates.
Highlights
- Isolated edits with durable follow-on work. Worktree isolation lets turns
edit in parallel against one checkout without spilling changes into other
turns. Because each turn syncs its edits back separately, code edits and Git
history changes must happen in separate turns. Durable continuations handle
that split automatically: one request can edit, sync, commit, and push, with
each step waiting for the previous step to succeed. - Persistent watches. Monitor changes through approved MCP read tools, with
progress visible in chat and the status panel. Separately authorized watches
can trigger bounded local agent work. - Shared agent tools. CLI and optional native MCP tools expose bounded
topic-history search and same-topic agent routes. Task-branch pushes and PRs
require an explicitly activated publication scope and a separate review. - More agent options. Adds the Antigravity (
agy) harness, improves Codex
and Claude sign-in, and supports model/effort selection with/. Adds remote
setup for Claude and Codex quota tracking. - Richer responses. Display Markdown tables and Mermaid charts and diagrams
in agent responses, with Mermaid support in Markdown previews too. - Better search. AND/OR keyword search and clearer result highlighting.
- Easier Python setup. The public installer can offer to install uv and a
compatible Python, including when run through the documented piped command,
without replacing system Python. - Reliability improvements. Better cancellation, queue recovery, reconnects,
context sync, scrolling, and attribution of per-turn stats and diffs.
Requirements
Python 3.10+ is required. Enable worktree.enabled for durable continuations.
Watches and automatic watch execution remain opt-in and require separate grants.
Antigravity requires native agy installation and sign-in.
Shore is experimental, disabled by default, and not available as a supported
production service in 0.1.6. This release publishes only the Python package;
no Shore browser client or bootstrap is deployed. General cross-topic routing
is not included. Release tags and GitHub releases are outside task-publication
and continuation scope.
v0.1.5 - WebSocket to Sync Multi-Devices
Release v0.1.5
curl -fsSL https://agentsquid.ai/install.sh | bash
Major
- Synchronized conversation across browser tabs and mobile devices via WebSocket. (Moved from SSE)
- Cross-tab consistency: Pinned items, active work, queue state, and usage gauges stay synchronized when Squid is open in multiple tabs or devices.
- Improved local-model support: Ollama management is faster and more dependable, additional local providers such as OMLX are easier to configure, and provider links can be customized.
- Topic subtree filtering: Selecting a topic can now include its dot-separated subtopics, making related conversations easier to find together.
Minor
- Access to TailScale domain via HTTPs & IP:PORT via HTTP
- Coding agent sign-in within AgentSquid
- Clearer Claude automatic setup: Reading Claude credentials from Chrome requires opening Squid directly at 127.0.0.1 or localhost. Squid now explains this limitation when accessed through Tailscale.
- More accurate usage information: Failed or cancelled requests no longer show misleading token, duration, or quota values. Usage gauges also stay synchronized across tabs and devices.
- Mobile improvements: Fixed truncated live commands, authentication input problems, and several layout and navigation issues.
- General polish and stability: Improved error messages, recovered-response presentation, quota readability, status wrapping, highlighting, and handling of unavailable statistics. Extensive realtime and browser test coverage was added as well.
v0.1.4 - Ollama as a provider!
v0.1.3
Release v0.1.3
curl -fsSL https://agentsquid.ai/install.sh | bash
Added
- Added a version update check button.
- Polished the reply button UI.
Fixed
- Resolved bugs in the version update logic.
- Fixed issue where the delete button was hidden on long auto-complete prompts.
v0.1.2
v0.1.2
Added
- Reply button route updating (#topic@agent) and cross-session context pinning.
- Hierarchical topic routing (#topic.sub1.sub2@agent) with shared parent topic memory.
Fixed
- Postponed worktree directory copying until prompt execution starts rather than queue time.
- Corrected empty stats filters for agents/topics on initial installation.
- Resolved version update failure.
v0.1.1
Squid v0.1.1
AgentSquid is free, open-source software that brings Claude Code, OpenAI Codex, Cursor Agent, OpenCode, and Pi into one shared workspace. Run parallel prompts, control live processes, search history across models, track per-prompt tokens and quota, and reach your local repos and filesystem from mobile.
Quick Start
- Install at least one supported CLI and check whether you can use it (/login):
| Backend | CLI | Install |
|---|---|---|
| Claude Code | claude |
curl -fsSL https://claude.ai/install.sh | bash |
| OpenAI Codex | codex |
curl -fsSL https://chatgpt.com/codex/install.sh | sh |
| Cursor Agent | cursor-agent |
curl -fsS https://cursor.com/install | bash |
| OpenCode | opencode |
curl -fsSL https://opencode.ai/install | bash |
| Pi | pi |
curl -fsSL https://pi.dev/install.sh | sh |
- Install and run AgentSquid!
curl -fsSL https://agentsquid.ai/install.sh | bash- Open in your browser:
http://127.0.0.1:8000
v0.1.1
Highlights
- Published
agentsquidfor PyPI/pipx installation and added release workflows for PyPI and TestPyPI. - Added update checking against PyPI releases, including an upgrade-on-restart flow for pipx installs.
- Improved Squid worktree handling, including dirty worktree tracking, recovery behavior, revert handling, and admission race fixes.
- Added AgentSquid version display in settings.
- Improved Codex image handling by persisting generated images.
- Refined chat/composer behavior, including streamed whitespace preservation, queued prompt cancellation, attached file clearing, and duplicate edit tool record cleanup.
- Fixed Claude
tool_useevents being emitted twice per turn. - Updated README install and quick start documentation.
- Added and expanded e2e/unit coverage across route chains, diff UI, stop behavior, update notices, runners, server, stats DB, and worktrees.
Bug Fixes
- 'Restart Server' function is fixed which was caused by changing distribution from GitHub to PyPI.
Release
- Final version bump from
0.1.1rc7to0.1.1. - Tag:
v0.1.1 - Commit:
b64a6e9
Squid v0.1 — Initial release
Squid v0.1
Agent-Squid is a meta harness tool which lets you run Claude Code, OpenAI Codex, Cursor Agent, OpenCode, Pi coding agents locally with multiple providers, such as, Claude, GPT, Cursor, OpenCode, NVIDIA, DeepSeek, local LLMs via one browser UI. Your agents still run on your own machine, with your repo, shell tools, credentials, and native CLI sessions. Squid gives those sessions names, global search, history, queues, controls, and phone/tablet access.
Quick Start
- Install at least one supported CLI and check whether you can use it (/login):
| Backend | CLI | Install |
|---|---|---|
| Claude Code | claude |
curl -fsSL https://claude.ai/install.sh | bash |
| OpenAI Codex | codex |
curl -fsSL https://chatgpt.com/codex/install.sh | sh |
| Cursor Agent | cursor-agent |
curl -fsS https://cursor.com/install | bash |
| OpenCode | opencode |
curl -fsSL https://opencode.ai/install | bash |
| Pi | pi |
curl -fsSL https://pi.dev/install.sh | sh |
- Install and run AgentSquid!
curl -fsSL https://agentsquid.ai/install.sh | bash- Open in your browser:
http://127.0.0.1:8000
Create agents in the UI. An agent is a named config:
agent name (@agent) = harness (Claude Code) + provider (Claude) + model (claude-sonnet-5) + working directory (CLAUDE.md, etc.)
Supported harnesses and providers
Squid separates the coded CLI integration from the model endpoint/account:
- Harness — the local coding-agent CLI Squid drives: Claude Code, OpenAI Codex, Cursor Agent, OpenCode, or Pi.
- Provider — the account or endpoint behind the harness: Claude, GPT, Cursor, OpenCode, NVIDIA NIM, DeepSeek, or a local models supporting standard APIs.
- Agent — your named lane in the UI:
name = harness + provider + model + model working directory.
Provider configuration lives under providers: in ~/.squid/squid.yaml. Native subscription providers work with zero config where the CLI is already logged in; API-key providers such as NVIDIA NIM, DeepSeek, and local OpenAI-compatible endpoints are configured once and can be reused by compatible harnesses.
Highlights
- Named topic and agent lanes with durable sessions:
#topic@agent - Parallel fresh-context prompts:
#topic@agent! - Limited lookback for goldfish mode:
#topic@agent!3 - Side-by-side work across Claude Code, Codex, Cursor Agent, OpenCode, Pi, and compatible custom provider endpoints
- Context pins for using selected responses as explicit context for other topics, agents and sessions.
- Live response progress, queue state, status output, and tool-call display
- Per-prompt input, output, cache, cost, duration, and quota signals where the harness/provider exposes them
- Provider-scoped real-time quota gauges, including static gauges for free-tier or local providers
- Topic-scoped Git diff tracking through
code_rootsshared across agents. - Queue and process controls: stop, stop-all, drain, clear, compact, and reset
- Global, topical, by route (#topic@agent+) search
- Topics view with searchable lanes, topic actions, and browser back navigation after opening a topic
- Browser UI on desktop, phone, and tablet via optional Tailscale HTTPS access with
/remoteQR code - Persistent user configuration, database, context, and logs under
~/.squid/
See the README for command syntax, agent creation, provider configuration, remote access, and security details.
