Pochi

Type: full-code · Vendor: TabbyML · Language: TypeScript · License: Apache-2.0 · Status: active · Status in practice: emerging

Links: homepage docs repo

Open-source AI coding agent shipped primarily as a VS Code extension by TabbyML: reads/writes files, runs commands, supports MCP, isolates concurrent tasks in git worktrees, and accepts custom/self-hosted LLMs.

Description. Pochi is TabbyML's coding agent. It lives in VS Code (Marketplace and OpenVSX), with global config at ~/.pochi/config.jsonc and per-project overrides at .pochi/config.jsonc. The agent uses file-read/write/edit/exec tools to handle multi-step development tasks. MCP support lets it integrate APIs, databases, and internal services. Parallel Agents isolate concurrent tasks in their own git worktrees, so multiple tasks stay live without stashing. Pochi accepts any LLM provider via API key and explicitly targets self-hosted/fine-tuned models.

Agent loop shape. Tool-using agent loop inside VS Code: the agent reads project files, executes shell commands, applies edits, and reaches external systems via MCP. Each top-level task can be run as a Parallel Agent in its own git worktree, so independent tasks have isolated workspaces and do not collide on file state.

Primary use cases

VS Code coding agent backed by any LLM provider
self-hosted / fine-tuned model deployments
parallel git-worktree-isolated tasks
MCP-extended internal workflows
GitHub-issue-to-PR pipelines from the sidebar

flowchart TD user[User task in VS Code] --> config[Layered config global plus project] config --> loop[Agent loop] loop --> propose[LLM proposes tool] propose --> tool{Tool} tool --> read[Read files] tool --> write[Write files] tool --> exec[Execute command] tool --> mcp[MCP server] read --> obs[Observation] write --> obs exec --> obs mcp --> obs obs --> loop loop --> parallel{Parallel Agent?} parallel -->|yes| worktree[Isolated git worktree] parallel -->|no| ws[Workspace] worktree --> done[Done] ws --> done

Key concepts

Tool use → tool-use (docs) — File reads/writes, command execution, codebase search.
Custom / self-hosted models → multi-model-routing (docs) — Any LLM provider via API key; self-hosted and fine-tuned models supported.
Parallel Agents → subagent-isolation (docs) — Each task runs in its own git worktree; multiple tasks active simultaneously.
MCP → mcp (docs) — Model Context Protocol integrates APIs, databases, internal tools.
Layered config (docs) — Global ~/.pochi/config.jsonc plus per-project .pochi/config.jsonc.