Lykos
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
Lykos
Status: under active development. Lykos is pre-release (0.0.1). Features are incomplete, commands and configuration may change without notice, and some providers are untested against live accounts. Use it on code you have committed, and expect rough edges.
Lykos is a coding agent for the terminal, built on the Pi runtime. It is built to be delegated to: an orchestrator hands it a task, and Lykos does the work behind an approval gate, records every model call, and reports what it verified.
Made for orchestrator–worker workflows. Over MCP, an orchestrator starts sessions without waiting (
lykos_start), checks on them (lykos_status), steers or stops them, and gets a result that separates what was verified from what is only the agent's claim.Fast to answer. A warm pool of session processes serves MCP calls without a cold start, and rarely used tools stay out of the prompt until searched.
Efficient with tokens. The prompt prefix stays byte-identical for the whole session so provider caches keep hitting, and long sessions are pruned and compacted in place.
Accountable. Every call lands in a SQLite ledger with its tokens, cache reads and writes, and cost, so spend is a query rather than a guess.
Customizable. Specialists, skills, extensions, permission rules and models are set in YAML, per user or per project.
It runs three ways:
Standalone, as a full-screen terminal UI or a one-shot task runner.
As an MCP worker under an orchestrating agent, which calls Lykos specialists as its own tools.
Under human supervision through a desktop app. Not built yet.
Features
An approval gate by default. Three permission modes (
manual,accept-edits,auto), remembered per project, plus path rules where deny always wins. Secret files and.git/are protected out of the box.Specialists. Named bundles of model, tools, skills, prompt and permissions:
plan,build,explore,general,review,test-writer,implementer, or your own in YAML. Switch mid-session.A full tool set. Hash-anchored
edit,grep,glob,ast_grep/ast_edit, anlspclient, background jobs (hub),web_search(keyless by default) andweb_fetch,github,eval, memory tools andsecurity_scan. Rarely used tools stay out of the prompt until searched.A ledger. Every model call is recorded in SQLite with its tokens, cache reads and writes, and cost.
lykos show,lykos costandlykos toolsread it.Nothing is deleted. Every session is archived and resumable. Checkpoints let you rewind the working tree to any turn without touching your branch.
Sealed TDD. A test writer's failing tests are signed; an implementer cannot change them, only dispute them.
Questions go up, not out. A delegated session asks its coordinator, not a terminal nobody is watching, and the coordinator answers or rules on disputes.
Model access. Pools of API keys and OAuth logins per provider, with rotation and cooldown when a credential hits its limit.
Optional containment. Run a session's commands inside Docker or Podman. Off unless you ask.
Memory. The agent may propose a lesson; only you accept it.
Related MCP server: deepseek_harness
Requirements
Node.js 22 or newer
git
Docker or Podman, only if you want containment
Lykos is developed on Windows. Other platforms are less tested.
Install
Lykos is not on npm yet. Install from source:
git clone https://github.com/rakun68/Lykos.git
cd Lykos
npm install
npm run build
npm linknpm link puts the lykos command on your PATH. Check the install:
lykos --versionQuick start
In the project you want to work on:
lykos initThis creates .lykos/config.yaml. The default model is DeepSeek. Give Lykos a key, either through
the DEEPSEEK_API_KEY environment variable or its credential store (the key is read from stdin):
lykos login deepseek --api-keyFor OAuth providers, run lykos login <provider>. Then check the setup:
lykos doctorStart the terminal UI:
lykosOr run one task and exit:
lykos run "add a --verbose flag to the CLI and a test for it"A task run this way works on a branch and worktree of its own, so your working tree is not touched.
Common commands
Command | What it does |
| Start the terminal UI |
| Run one task ( |
| Continue an earlier session, or pick one |
| Replay a session: edits, checks, checkpoints, cost |
| Restore the working tree to a turn checkpoint |
| Spend from the ledger |
| Manage credentials and see usage windows |
| List the specialists available here |
| Review proposed and accepted memories |
| Serve Lykos as an MCP worker over stdio |
| Every command and flag |
Use Lykos as an MCP worker
Any MCP client can run lykos mcp over stdio. To register Lykos for a project:
lykos export claudeThis adds lykos to the project's .mcp.json and writes one agent file per Lykos specialist. The
orchestrator can then call lykos_<specialist> tools, or start sessions in the background with
lykos_start and follow them with lykos_status. The contract is in
docs/reference/mcp.md.
Configuration
Configuration is layered: shipped defaults, then ~/.lykos/config.yaml, then the project's
.lykos/config.yaml, then the specialist, environment variables and flags. Runtime state lives in
.lykos/; lykos init adds the ledger, sessions and cache to your .gitignore. The keys are in
docs/reference/cli-and-config.md.
Documentation
docs/reference/: CLI and configuration, the terminal UI, sessions and archive, specialists, model access, the MCP contract, the ledger
docs/adr/: architecture decisions and why they were made
GLOSSARY.md: the project's terms
AGENTS.md: orientation for contributors and coding agents
Coming next
Rust where it counts. Rewriting hot-path modules in Rust for startup time and speed.
Detaching from Pi. Lykos becomes its own fork of Pi rather than depending on it. It runs alongside an existing Pi install and no longer needs a
.pi/directory.Lykos orchestrating Lykos. Running a task, then a whole phase of tickets, through review cycles from the CLI.
A rebuilt terminal UI with a command palette, model picker and inline diffs, then a desktop supervisor.
License
MIT. Some files contain copied MIT-licensed code: see docs/copied-code.md.
This server cannot be deployed
Maintenance
Related MCP Connectors
Hosted coding agent: start, follow up on and stop tasks that can open GitHub pull requests.
Persistent memory, hybrid search and a goal graph for AI agents, over stdio or remote HTTP.
MCP server (stdio): check if URLs are alive (status, redirects, latency) via the AgentForge API
Coordination hub for AI coding agents: message teammates, ask humans, audit every event.
Related MCP Servers
- FlicenseNot gradedqualityCmaintenanceProvides safe repository operations, skill reading, command execution, and test detection for coding agents over stdio.-
- AlicenseAqualityBmaintenanceA task-level STDIO MCP server that lets Codex or any other MCP client hand off scoped coding jobs to an asynchronous worker agent which reads the code, edits files, and runs tests, while the client keeps ownership of planning and acceptance. Exposes submit, wait, query, follow-up, and cancel tools so multiple clients can queue and monitor tasks against a chosen project root.51MIT
- AlicenseNot gradedqualityFmaintenanceEnables a primary coding agent to delegate bounded implementation and repair tasks to a local OpenAI-compatible model that edits a copy of the project, runs configured checks, and returns artifacts for review. Exposes stdio tools for capability discovery, task submission, progress polling, artifact reading, and cancellation.MIT
- AlicenseNot gradedqualityAmaintenanceEnables compatible agents to call registered local tools and read resources through one stdio server, including harness preparation, knowledge search and document management, memory retention/recall, correction proposals, advisory decisions, tool/skill discovery, and log/source inspection.2Apache 2.0