Nausicaa
A recoverable multi-agent harness that reshapes its Lane topology around each task.
Per-dimension scores and reasoning
The evidence shows host-managed authority, enforceable per-member tool narrowing, explicit MCP grants, a read-only Plan boundary, effect/scope metadata, and tests that treat Skill text as untrusted data. Data projections, Teto visibility, credential location, stopping, and recovery are described concretely. Deductions apply because commands and writes inherit the current user's authority by default and per-operation confirmation is not established; credential deletion and disclosure guidance exist, but encryption at rest, file modes, and rotation are unspecified; dependencies are pinned and installed in CI, but the supplied material does not show vulnerability scanning, supply-chain signing, or lockfile review; and checkpoints, cloning, and recovery do not provide transactional rollback for file or external effects. Attribution to the author, repository, upstream influences, and third-party notices is thorough, while publisher identity remains unknown as required.
The README, package metadata, workflows, and offline tests agree on the version, Node requirement, check commands, Lane/Teto behavior, and recovery boundaries. Tests also establish specific states or failure codes for exhausted budgets, unknown cost, tampered manifests, unsafe paths, and missing resume boundaries, supporting full scores for consistency and failure messaging. Dependency availability is reduced because provider, daemon, RPC, and edge integrations are expressly evolving, model access depends on external services, and comprehensive retry or degradation behavior is not shown.
The material addresses interactive TUI, one-shot and JSON operation, read-only daemon attachment, multiple providers and models, Skills, MCP, Workers, Teams, and cross-Run A2A. Tool effects, scopes, nested-team depth, member capabilities, Plan-mode restrictions, and disable switches give strong capability boundaries. Trigger precision is reduced because Teto's unsolicited “high-value” threshold and task-to-Skill matching remain model judgments rather than fully specified rules. Environment evidence centers on Node 22.19+/24 and Ubuntu CI, without comprehensive proof for other operating systems, terminals, or sandbox backends.
The bilingual README is well organized and supplies terminology, command tables, installation, development steps, and concrete topology examples. Lane, Teto, Team, Worker, and addressing names are consistently defined. Limitations are unusually explicit: beta status, no hostile-code security boundary, no generic MCP OAuth, latest-beta-only support, and evolving integrations. The MIT text matches package metadata. Deductions apply because there is no standalone FAQ and some detail is delegated to files not supplied; CHANGELOG, Releases, and update commands are referenced without the actual release history or compatibility policy; and while an author and reporting paths exist, long-term ownership, response targets, and multi-maintainer governance are not defined. Registry verification of the publisher is absent but is treated as unknown, not suspicious.
The CLI, TUI, JSON events, import/export, topology inspection, recovery, collaboration messages, and structured failure records produce outputs usable in development and automation. Dynamic Lanes, an observer, resumable Teams, and host-owned facts offer clear marginal value over a fixed linear agent loop. Cost-benefit is reduced because auxiliary Lanes add model calls and tokens, while members intentionally lack default token, call-count, and duration budgets. Tests cover shared meters, request limits, stopping on uncertain cost, and a bounded Teto overhead case, but ordinary-run cost forecasting, default ceilings, and broader benefit benchmarks are not shown.
Core claims trace to package scripts, CI and publishing workflows, and concrete tests for permissions, edge tools, recovery, cost controls, provenance hashes, and tamper rejection. README, package metadata, LICENSE, SECURITY, workflows, and tests corroborate one another. The sources also separate claims from caveats by labeling the product beta, limiting claims to offline coverage, stating that local configuration does not verify remote access, and identifying evolving integrations. Full scores here reflect strong static traceability in the supplied files only; no execution or independent validation was performed.
- Commands and file writes use the current user's authority by default. Run untrusted repositories or code inside an independently managed sandbox or VM, and narrow permissions with /permissions first.
- Credentials are stored in ~/.nausicaa/credentials.json, but encryption, file-mode enforcement, and rotation are not documented. Verify local access controls and keep credentials, private code, and logs out of issues, pull requests, and fixtures.
- Teto is enabled by default, and Team members inherit the lead's authorized workspace tool catalog unless narrowed. For sensitive work, review projected data, restrict tools per member, and use --main-only, --no-worker, or disabled edge sources where appropriate.
- Members have no default token, request-count, or duration budgets, and asynchronous work can continue while an interactive session remains open. Use provider limits, external monitoring, and explicit stop policies to control cost and effects.
- This is version 0.1.9 beta, with only the latest beta on the default branch maintained; provider, daemon, RPC, and edge integrations are still evolving. Review changes and validate each upgrade in the target environment.
What does this agent do, and when should you use it?
Nausicaa is a general-purpose command-line agent harness that models a Run as a changing topology of Lanes instead of a fixed linear loop. Its root agent, the Teto observer, delegated Workers, Team members, and cross-Run A2A links can be composed as work evolves, with each Lane retaining its own address, context, lifecycle, and recovery state. The model chooses when and how to collaborate, while the Host enforces identity, authority, durability, and recovery. It supports an interactive TUI, one-shot `--print` runs, `--json` event output, a daemon, read-only attachment, session recovery, imports and exports, Skills, permissions, and MCP configuration. It requires Node.js 22.19.0 or newer and documents access through OpenAI API, ChatGPT subscriptions, Anthropic, OpenRouter, and other supported providers; version 0.1.9 is labeled beta, with provider, daemon, RPC, and edge integrations still evolving.
After the user starts nausicaa inside a project directory, the root nausicaa Lane accepts the task and can run commands or edit files with the current user's permissions, subject to restrictions applied through /permissions. It can delegate work to Workers, incrementally create Teams with team_create and team_assign, coordinate group work through team_message, send private A2A messages, and communicate with other Runs. task_wait waits for results while yielding when new collaboration events arrive, and team_status returns an immediate snapshot; final member reports enter the shared task thread for synthesis and acceptance by the creating Lane. Teto observes public projections of messages, tool requests, and bounded terminal states after completed steps, looking for intent drift or a better next move and sending unsolicited advice only when it is new and consequential. The Host preserves Lane identity, authorization, sessions, and recovery data, exposed through commands such as /resume, /session, /clone, /export, /import, --continue, and --resume <run-id>. Built-in Skills cover codebase-map, task-plan, and code-review, while project Skills can be discovered from .agents/skills, .pi/skills, and skills directories.
- An engineering lead investigating a large repository can give researcher and reviewer Lanes separate contexts, run their work concurrently, and retain final synthesis and acceptance.
- A developer running a long coding or analysis task can interrupt the current response without discarding asynchronous Team work, then resume from durable member reports.
- A solo developer concerned about requirement drift can keep Teto enabled as a low-interruption observer that evaluates work at completed-step boundaries.
- A team evaluating several model providers can manage OpenAI API, ChatGPT subscription, Anthropic, OpenRouter, and other supported logins and models from one TUI.
- An engineering group integrating agent runs into scripts can use one-shot output, JSON events, daemon operation, read-only attachment, and explicit Run recovery.
- A repository maintainer can add project-specific Skills and configure authorized HTTP or stdio MCP services to extend the working environment.
What are this agent's strengths and limitations?
- Lanes are independently addressable and resumable, allowing Teams, Workers, observers, and cross-Run A2A connections to be composed dynamically instead of following one fixed loop.
- Teto provides a separate observation path, evaluates only after complete steps, and deduplicates identical unsolicited notes to limit disruption.
- Identity, enforced permissions, durability, and recovery belong to the Host; a Team lead can narrow member tools and control whether members may create nested Teams.
- Interactive TUI, one-shot output, JSON events, daemon execution, and read-only attachment support both hands-on work and automation.
- Documented multi-provider login and per-session model selection reduce dependence on a single model API.
- Version 0.1.9 is explicitly beta, and provider, daemon, RPC, and edge integrations are still evolving, creating upgrade and operational risk.
- Adoption requires Node.js >=22.19.0, npm installation, and credentials for an external model provider.
- Commands and file writes inherit the current user's permissions by default, so adopters must configure
/permissionswhen a narrower execution boundary is required. - Generic MCP OAuth is unavailable in this release, and any MCP configuration change requires a restart.
- Teto sees only public projections and bounded tool states rather than the owner's complete context or raw tool results, limiting what it can reliably diagnose.
- Nested Teams are capped at three levels, which may require flattening more elaborate delegation structures.
How do you install or deploy this agent?
Install Node.js >=22.19.0 and prepare credentials for at least one supported model provider. Install the package globally and start it in the target project:
npm install -g nausicaa-harness
cd /path/to/projectnausicaa
For source development:
git clone https://github.com/jackispm/nausicaa-harness.git
cd nausicaa-harness
npm ci
npm run build
npm linkThe documented checks are npm run typecheck, npm test, npm run eval, and npm run test:smoke.
How do you use this agent?
On the first TUI launch, run /login and choose a provider and one of its supported authentication methods; OpenAI API access and ChatGPT subscription access are separate entries. Documented shell alternatives include nausicaa auth login openai api-key, nausicaa auth login openai-codex oauth, and nausicaa auth login openrouter api-key. After authentication, use /model to select the session model and /thinking [level|default] to choose a supported reasoning level. Enter the task from the project directory, inspect the active topology with /list-agents, constrain access with /permissions, and recover work with /resume or --resume <run-id>. Use nausicaa --print for a one-shot answer, --json for events, --main-only when creating a Run without Teto, and --no-worker to disable Worker delegation. Skills and MCP are optional and are managed through /skills and /mcp; MCP configuration changes require a restart.
How does this agent compare with similar options?
Nausicaa uses pi-ai for provider transport and cites Pi coding-agent, Prime Agent, and DeepSeek Harness as references for selected boundaries. It identifies Teto, dynamic Lane topology, and Host-owned durability as its own runtime contracts. The supplied material includes no feature matrix, benchmark, or migration study against those projects, so it does not support a broader claim of superiority.
FAQ
Does Nausicaa require a particular paid model service?
Can it modify my repository by default?
/permissions can restrict them. Team members normally inherit the lead's Host-authorized workspace tool catalog, which the lead can narrow further.What happens to Team work if I interrupt a response?
Esc or Ctrl+C interrupts the current answer while preserving session context and asynchronous Teams. Member reports remain available. /stop, session closure, or task cancellation stops the corresponding work.