Yagent

A local-first AI agent for coding, audits, reviews, web search, and research.

Yagent is written in Go and runs against OpenAI-compatible inference servers: Ollama, llama.cpp, or a cloud endpoint you explicitly configure. It owns the agent loop, memory, orchestration, and tools—no LLM framework required.

Why Yagent

• Make changes safely. Stream tool use, review risk-gated writes and diffs, approve fs_patch hunks individually, and cancel a running turn with Esc without losing the session. A loop guard stops repeated generation; /yolo is available when you deliberately want automatic approvals.
• Verify work before calling it done. Built-in diagnostics, targeted tests, runtime smoke checks, syntax/YAML/JSON validation, and semantic-diff protection catch common mistakes. Goal and test gates can refuse completion until declared checks pass; codegen mode is tuned for greenfield builds.
• Keep every turn recoverable. In Git repositories, turn commits preserve pre-existing work and power crash-safe /undo; /diff shows the cumulative session change before you keep it. SQLite sessions support resume, search, and Markdown or HTML export.
• Work across longer tasks. Context is token-budgeted with old tool output pruned before summaries; hybrid memory combines vector search, FTS5, importance, and recency. Skills provide reusable SKILL.md procedures with progressive disclosure and optional verification.
• Understand a codebase before editing it. A gitignore-aware, tree-sitter index supports structural search, surgical symbol reads, call references, impact analysis, topology, unused-symbol candidates, and environment audits. Empty directories are first-class greenfield workspaces: Yagent detects available toolchains, asks for a stack or offers a scaffold, then enables project verification as soon as a manifest exists.
• Delegate and automate deliberately. Use goal mode, resumable checkpoints, declarative playbooks, parallel read-only subagents, and a shared scratchpad. clarify and plan provide structured handoffs; /plan keeps exploration read-only.
• Stay local by default, extend when needed. Web results are treated as untrusted data; DuckDuckGo, Mojeek, and SearXNG are supported. MCP tools, deterministic hooks, an optional advisor, and a separate summarizer model extend the workflow without changing the core loop.
• Choose the model and interface that suit the job. The Bubble Tea TUI and plain REPL share one runtime. The TUI includes provider/model selection, settings, sessions, skills, tool activity, workspace overview, transcript search, themes, accessibility modes, and notifications. Local models are discovered live; cloud choices come from models.dev.

For the full command and safety model, see the tool documentation. For local-model results and recommended settings, see the benchmark guide.

Install and run

go install github.com/Mechres/Yagent@latest

Requires Go 1.22+ (built and tested with Go 1.25). Tree-sitter indexing requires cgo, so install a C toolchain too.

Start a local inference server and pull a chat model plus an embedding model:

ollama serve                 # or: llama.cpp llama-server --embeddings
ollama pull qwen3vl:8b       # or load a GGUF with llama.cpp
ollama pull nomic-embed-text

Yagent defaults to Ollama at http://localhost:11434. Configure another OpenAI-compatible endpoint through YAGENT_SERVER_URL, YAGENT_MODEL, or config.yaml; a repository-local .yagent/config.yaml overrides the global configuration. See config.example.yaml for every setting.

Quickstart

yagent init                                      # write a starter configuration
yagent doctor                                    # verify server, model, embeddings, and toolchain
yagent chat                                      # open the streaming TUI (--plain for the REPL)
yagent chat --goal "refactor the parser package"  # run an autonomous goal loop
yagent chat --goal "build a tetris" --check "tetris.cpp exists"
yagent chat --playbook release-checklist         # run a declarative workflow
yagent bench --repeat 3                          # measure and record a model baseline
yagent calibrate                                 # find a sampling recipe for the current model

Useful follow-ons:

yagent sessions export <id> --format html                 # share a session
yagent export-dataset --format sharegpt --output data.jsonl  # export verified trajectories
yagent export-dataset --format dpo --output preferences.jsonl

In the TUI, /model selects a provider and model (local models are auto-detected; cloud choices refresh from models.dev).

Useful TUI controls

Command / key	Effect
/tools	Browse tool calls; f filters, g jumps to transcript activity, Enter expands details, and PgUp/PgDn/Home/End navigate.
/workspace	Show workspace, branch, context use, tool count, undo availability, and queued-work state. A compact drawer appears during active turns on wide terminals.
/sessions <query>	Filter sessions by ID or generated title. In the browser, p previews, s changes ordering, n renames, and * pins.
Enter while working	Queue one follow-up message; a later queued message replaces the earlier one.
a / x during patch review	Accept all remaining hunks / reject all remaining hunks.
/set ui.accessibility high-contrast	Persist a high-contrast TUI palette; set standard to restore it.
/set ui.accessibility ascii	Use ASCII labels instead of emoji for limited terminal fonts. NO_COLOR=1 suppresses color styling.
/set ui.reduced_motion true	Static spinner (no animation) for vestibular sensitivity.

Security & privacy

• Local-first by default — LLM and embedding requests go only to the configured server. By default that's a local Ollama/llama.cpp — nothing leaves the machine.
• Opt-in cloud — set api_key (or YAGENT_API_KEY) and pick a cloud provider via /model (OpenCode Zen/Go, DeepSeek, OpenRouter, Groq, Together, Mistral, NVIDIA NIM) to run the whole loop in the cloud; consult has its own api_key for a separate advisor model. Keys entered via the TUI /model prompt or REPL /key are stored in the config file's api_key field (/key clear removes them); keys from environment variables are never written to disk and take precedence. Both are deliberate opt-ins — the default config stays local.
• Redaction — before anything is written to SQLite (messages, summaries, memories) or exported, likely secrets (api_key/token/password/bearer values) and home paths are scrubbed to [redacted]/[home] markers. This is a heuristic guard, not a security boundary. Session exports warn when they contain these markers.
• Approvals + sandbox — write/destructive tools require explicit approval (unless /yolo, which is pre-granted consent; approved tool+args auto-approve for the rest of the session); shell.sandbox: bwrap additionally wraps shell_exec in bubblewrap and fails loudly if bubblewrap isn't installed. Read-only plan mode (/plan) lets you keep the agent in explore-only mode until you approve a plan.
• Untrusted web content — web_fetch/web_search results are wrapped as <untrusted data ...> data (never commands), closing prompt-injection via a fetched page.
• No telemetry — nothing leaves your machine except explicit web_search/web_fetch calls and, if configured, the consult advisor.

Documentation

Doc	Contents
AGENTS.md	Build/test commands, conventions, constraints — read this first
docs/PLAN.md	Milestones M1–M7 with tasks and acceptance criteria
docs/design/architecture.md	System design, module layout, decision log
docs/design/agent-loop.md	Agent loop, tool calling, context budgeting
docs/design/memory.md	Memory layers, storage schema, retrieval
docs/design/skills.md	Hermes-style skills: procedural memory, SKILL.md format, approval gate
docs/design/tools.md	Tool specifications and safety model
docs/models.md	Model quirks from acceptance runs (tool-call reliability, embeddings)
docs/models-benchmark.md	Which model to run, what to expect, and recommended settings (benchmarked on an RX 6700 XT)
docs/RESEARCH-other-tools.md	What we borrow from opencode/aider/plandex (and the broader field) and what we deliberately skip
config.example.yaml	Annotated configuration reference
CHANGELOG.md	Release history

Development

make build     # or: go build ./cmd/yagent
make test      # go test ./...
make vet       # go vet ./...
make race      # go test -race ./...

License

MIT — see CONTRIBUTING if you'd like to help.