Harness encyclopedia
The harnesses, running on a server.
TeamDay runs the agentic harnesses developers already trust — server-side, around the clock. Pick a harness per agent. Mix them on a single mission. New models from each vendor light up the moment they ship — no integration sprint.
Anthropic
Claude Code
The harness for long-horizon, verifiable engineering work.
Anthropic's official agentic coding harness. Runs the Claude family — Fable 5, Opus 4.8, Sonnet 4.6. Strong on multi-hour missions, tool use, and self-verification.
Models
Read the encyclopedia entry
OpenAI
Codex
The harness for unified multimodal and computer-use missions.
OpenAI's agentic harness. Runs the GPT-5 family — including GPT-5.5 and GPT-5.5 Pro. First-class computer use, browser automation, and a unified text/image/audio/video architecture.
Models
Read the encyclopedia entry
Gemini CLI
The harness for huge-context reasoning and Google-stack work.
Google's agentic harness. Runs the Gemini family — Gemini 3.1 Pro, Gemini 2.5 Pro. Industry-leading 2M context window and tight integration with Google Workspace, Drive, and BigQuery.
Models
Read the encyclopedia entry
xAI
Grok Build
The harness for Grok-powered terminal work and fast model routing.
xAI's Grok Build CLI runs headless from TeamDay with streaming JSON events, session resume, and optional model selection through the xAI account or API key configured on the runner.
Models
Read the encyclopedia entry
Earendil Works
Pi
The harness for provider-agnostic CLI work through third-party tokens.
Pi is a minimal terminal coding harness with JSON mode, durable session IDs, and provider selection. TeamDay runs it with your Pi auth file or provider API-token environment.
Models
Read the encyclopedia entry
On the roadmap
Coming next
opencode
soonOpen-source, model-agnostic. Runs any LLM you point it at. Good fit for teams that want a fully open stack.
opencode.ai
Forge
soonProduction-shaped coding harness from forgecode.dev. Modeled around features, branches, and code review.
forgecode.dev
The roster grows whenever a credible new harness ships. Want yours added? Drop us a note.
Side-by-side
Capability matrix
What each harness actually ships, verified against the upstream source. Updated April 2026.
| Capability | Claude Code | Codex | Gemini CLI | opencode | Pi | Forge |
|---|---|---|---|---|---|---|
| Sub-agents | yes | yes (spawn_agent) | yes | yes | yes | yes (Forge/Sage/Muse) |
| Skills | yes | yes | yes | yes (SKILL.md) | yes | yes |
| MCP client | yes | yes (+ MCP server) | yes | yes | yes | yes |
| Hooks | rich lifecycle | notify only | richest lifecycle | rich plugin hooks | yes | event flag only |
| Slash / custom cmds | yes | yes | yes (TOML) | yes | yes | yes (:prefix) |
| Permissions / sandbox | fine-grained | OS-level (seatbelt/landlock) | Policy Engine | per-agent rulesets | gates + path protect | worktree sandbox |
| Plugins | marketplace | marketplace | extensions | plugin SDK | TS extensions | — |
| Multi-model | Anthropic + Bedrock + Vertex + Foundry | OpenAI + Bedrock + Ollama + custom | Gemini only | 12+ providers | 15+ providers | 8+ providers |
| Sessions | --resume, /branch, Remote | SQLite, ephemeral | checkpoints + git snap, /restore | LLM-summary compaction | tree-structured (branchable) | resume + :compact / :clone |
| Surfaces | CLI/VS Code/JetBrains/Desktop/Web/iOS | TUI/CLI/Web/Desktop/IDE/AppServer | TUI/VS Code/GH Actions | TUI/web/desktop/IDE | TUI/JSON/RPC/SDK | TUI/ZSH/CLI |
| Headless / SDK | Agent SDK (TS+Py) | codex exec + AppServer + SDKs | -p, JSONL, @google/gemini-cli-sdk | createOpencodeServer/Client (HTTP) | 4 modes incl. SDK | CLI only |
| License | proprietary engine | Apache 2.0 | Apache 2.0 | MIT | OSS | Apache 2.0 |
Sourced from each upstream repo (Claude Code, openai/codex, google-gemini/gemini-cli, sst/opencode, badlogic/pi-mono, antinomyhq/forge).
How harnesses fit in TeamDay
Each agent on the roster has a harness, a set of MCP servers, and a mission. The harness handles reasoning, planning, and tool calls. The MCP servers expose your data, your platforms, and visual models like Seedance 2.0 and gpt-image-2. The mission tells the agent what success looks like.