Agent-Body

A DeepSeek Harness plugin layer that loads task-relevant tools and adds memory, reflexes, and attributed recovery.

Stars
★ 18
Last updated
1mo ago
License
MIT
Primary language
TypeScript

At a glance

How it runs
CLIAgent plugin / skillFramework
Works with
Platform-specific
Cost
Free, no paid service needed
Setup effort
Medium · a few setup steps
You'll need
DeepSeek HarnessNode.js 22.19 or 24PowerShell 7 (documented Windows path)Shell / CLINetwork accessLocal filesystem
Typical use
DeepSeek Harness operators whose growing tool catalog should expose only schemas relevant to the current task.
Not a fit if
  • Teams that do not run DeepSeek Harness
  • Production teams requiring a stable third-party API or SLA
  • Users requiring Linux or macOS parity with Windows
Maintenance
Commits in the last 6 monthslast commit 2026-09-30latest release v0.1.1 (2026-09-23)

What does this agent do, and when should you use it?

Agent-Body is a plugin layer for DeepSeek Harness rather than a standalone, general-purpose agent framework. Its components include the organism kernel, the Cortex memory organ, the Zero-residence context engine, and optional organ plugins coordinated through a nerve bus, heartbeat, and deterministic routing rules. In its reproducible 48-command cold-start benchmark, average estimated tool-schema usage falls from 55,154 to 8,433 tokens, a 84.71% reduction limited strictly to tool schemas rather than the full prompt or bill. The runtime records outcomes, adjusts routing synapses, attributes failures before selecting remedies, and consolidates long-term memory with deterministic rules while idle. The core demo and checks run without a host, network, model, or API key, but mounted organs and the body_* interfaces require DeepSeek Harness; the project is early-stage and Windows-first.

At agent/pre-step, the runtime converts command text into a nerve impulse. body_nerve applies deterministic intent rules, selects organs and named capabilities, and publishes the result through organism/impulse. After execution, tools/result updates learned routing: successful delivery strengthens a command-class-to-organ synapse, while failure weakens it. body_heal attributes errors to tool_missing, arg_error, permission, timeout, network, not_found, conflict, or unknown before choosing a remedy; arg_error is never automatically retried, and a wound closes only after that organ later succeeds. The heartbeat writes directives, proprioception, and homeostasis warnings to bloodstream.json and broadcasts organism/heartbeat. During deep sleep, Cortex derives pitfall, playbook, hotspot, unresolved, and fact memory cards from recorded experience. Zero-residence exposes zr_compact, zr_recall, zr_ledger, and zr_fast to replace resident content with pointers, reconstruct masked payloads verbatim from session logs, measure context cost, and run long commands asynchronously. body_status, body_map, body_cell, body_pulse, and body_tokens expose runtime health, ownership, capability units, event history, and tool-schema token accounting.

  1. DeepSeek Harness operators whose growing tool catalog should expose only schemas relevant to the current task.
  2. Framework maintainers who need an offline, model-free reproduction of routing, failure attribution, and reflex execution.
  3. Local-agent users who want outcomes retained across sessions and consolidated into deterministic long-term memory while idle.
  4. Automation teams that need to distinguish argument, permission, timeout, network, and other failures before retrying.
  5. Plugin authors building custom DeepSeek Harness organs with defineOrgan declarations for capabilities, permissions, signals, and fallbacks.

How do you install or deploy this agent?

To verify the core without installing the host, use Git and Node.js 22.19 or 24. No npm install, model, or API key is required.

git clone https://github.com/1420079678-ctrl/agent-body && cd agent-body
npm run demo
npm run check

Operational use requires a working DeepSeek Harness installation. The supported command below installs the organism, Cortex, and Zero-residence packages from the latest GitHub Release; restart the harness afterward.

rel=https://github.com/1420079678-ctrl/agent-body/releases/latest/download
dsh plugin --profile web add \
  "$rel/dsh-external-dsh-organism-0.1.1.tgz" \
  "$rel/dsh-external-dsh-cortex-0.1.1.tgz" \
  "$rel/dsh-external-dsh-zero-residence-0.1.0.tgz"

If the GitHub asset redirect is blocked, the documentation provides a jsDelivr path:

rel=https://cdn.jsdelivr.net/gh/1420079678-ctrl/[email protected]/dist
dsh plugin --profile web add \
  "$rel/dsh-external-dsh-organism-0.1.1.tgz" \
  "$rel/dsh-external-dsh-cortex-0.1.1.tgz" \
  "$rel/dsh-external-dsh-zero-residence-0.1.0.tgz"

The packages are not published to the npm registry. Some organs require a compiler, browser, or external binaries and must be installed from source with their replay scripts. The supplied material does not document DeepSeek Harness model credentials, so host credential requirements cannot be established here.

How do you use this agent?

Run the end-to-end demo, repository gate, and schema benchmark locally:

npm run demo
npm run check
npm run bench
npm run bench:check

After installing the plugins and restarting DeepSeek Harness, inspect health and ownership, then route a command without executing it:

body_status
body_map
body_nerve action=send text="Crawl this site and extract structured data"

body_status reports organs, heartbeat, fatigue, and integrity; body_map shows capability ownership; body_nerve returns the organs and capabilities selected for the command. A capability excluded by the initial gate remains retrievable through body_call. Use npm run verify for repository structure, JSON, link, and secret-hygiene checks. npm run verify:organs runs organ regressions and explicitly reports SKIP when the host runtime is unavailable.

What are this agent's strengths and limitations?

Pros
  • Schema gating has a reproducible in-repository benchmark: 48 cold-start commands average 55,154 to 8,433 estimated tokens, with bench:check detecting drift.
  • Routing, attribution, reflex handling, and sleep-time consolidation use deterministic rules, allowing core verification without a model or API key.
  • Recovery is cause-specific: argument errors are not blindly retried, and a wound remains open until the affected organ later succeeds.
  • defineOrgan gives plugins an explicit contract for capabilities, permissions, signals, handled failures, and fallbacks, while the kernels avoid dependence on one organ.
  • Vitals, wounds, synapses, memory cards, and the pulse stream are represented as inspectable runtime data.
Limitations
  • Core operation is tied to DeepSeek Harness; without that host, users get the offline demonstration rather than a deployable complete agent.
  • The project is explicitly early-stage and Windows-first; Linux CI covers the zero-dependency core more thoroughly than the organ integrations.
  • Schema savings decline as history keeps organs hot: the documented live measurement is 58%, so 84.71% is not a steady-state guarantee.
  • Ten of 48 benchmark commands require a second body_call, and large organs are affected by a ten-capability-per-organ exposure cap.
  • Only five of the 24 plugin packages currently have offline regression suites; the rest rely mainly on replay scripts.
  • Packages are not yet published to npm, and some organs require host-side compilers, browsers, or external binaries.

How does this agent compare with similar options?

Key facts side by side with the most closely related agents.

Agent Source review Form / cost Stars Updated Language Full support on
Agent-Body This agent 58 · Major gaps CLIFree ★ 18 1mo ago TypeScript —
ANOLISA Agent OS Layer 65 · Some gaps CLIFree + model costs ★ 684 1d ago Rust —
Hive Agent Harness 65 · Some gaps CLIFree + model costs ★ 11k 3d ago Python OpenAI API · Claude API
Auto Company 59 · Major gaps CLIFree + model costs ★ 3.2k today Python Codex · Claude Code

How does FollowAgents rate this agent?

FollowAgents source review · FARS-2.1
Major gaps
58/ 100 5-point scale 2.9 / 5
Trust 13/29
Reliability 8/14
Adaptability 12/18
Convention 14/18
Effectiveness 7/13
Verifiability 4/8
Why each dimension lost points
Trust13 / 29 · 2.2/5

The code distinguishes permission kinds and risk levels, marks high-risk permissions as requiring confirmation, allows only read-only remedies to run automatically, routes side-effecting remedies for model judgment, and bars automatic retry for arg_error; this supports least privilege and confirmation (2). The core describes itself as having no I/O, network, or model calls, and the README distinguishes the offline demo from host operation with some data-flow detail (2); retention details for persistent memory, logs, and sensitive data are thin (sensitive data 1). Dependency availability is partly explained through the zero-dependency core and host runtime requirements, but no complete dependency audit is shown (1). Recovery may cause side effects, while the sample does not show a complete effects inventory, per-action confirmation, or recovery mechanism (external effects 1, rollback 1). Copyright and repository location are stated, but publisher identity is unverified and does not establish maintainer identity (source attribution 1).

Reliability8 / 14 · 2.9/5

The README, benchmark claims, and CI workflow describe offline checks and baseline gates. However, package version, displayed README version, and pinned install versions differ, and counts such as 26 identities versus 24 packages require readers to reconcile stated scopes, reducing self-consistency (1). Node requirements, host prerequisites, the offline core demo, and the host runtime requirement for organ regressions are documented (dependency availability 2). Contract validation reports field paths; recovery classifies causes and explicitly rejects retrying argument errors, providing useful diagnostics (failure messages 2).

Adaptability12 / 18 · 3.3/5

Documentation addresses DeepSeek Harness plugin authors and operators, with Windows-first, offline-demo, host-install, and organ-authoring scenarios (2). It declares capabilities, permissions, fallback organs, and the host boundary, and warns that plugins run with host privileges; the sample does not establish every plugin's actual permission boundary (2). Routing rules, reflex conditions, and trigger events are described deterministically, though the supplied source excerpt does not show the entire routing implementation (2). Node versions, Windows-first support, host integration, and an independent offline core path are stated (2).

Convention14 / 18 · 3.9/5

The README provides contents, architecture, organ catalog, quick start, development guidance, FAQ, limitations, and evidence links; source and CI are also organized into core, plugins, and benchmarks (information architecture 3). Installation covers GitHub, CDN, and offline packages and explains restart and the unavailable npm path; actual use still depends on the external host and download sources (2). Organ, capability, and plugin names have contracts and version fields, but repository-wide version references are not fully aligned (2). Examples, FAQ, offline verification, and installation guidance are present (3). Early-stage status, host dependency, cold-start versus history trade-offs, second-hop retrieval, and environment constraints are disclosed (known limitations 3). The MIT license is complete and explicit (3). No verifiable changelog is in the sample, and version references differ, so version history is only partly supported (1). The README calls this a one-person project and gives an issue path, but the sample does not verify the person's identity (maintenance responsibility 2).

Effectiveness7 / 13 · 2.7/5

Tool calls, status, memory, recovery, and schema gating have clear user-facing entry points, and the offline demo and report help inspection; this sample cannot establish the full host experience (2). The stated aim is to reduce exposed tool schemas while providing deterministic routing, recovery, and memory, with benchmark figures and scope that indicate a concrete potential benefit (2). Cold-start savings, lower savings with history, and second-hop cases are disclosed, but the sample does not adequately quantify deployment, maintenance, or side-effect costs (1).

Verifiability4 / 8 · 2.5/5

The README points key figures to a benchmark report, corpus, and check command, and points demo claims to captured data; claims have traceable evidence paths, though the report itself is not included in this review packet (2). README, CI, package source, and security policy corroborate some claims, but the source sample is limited and several measured claims appear only in the README (1). The docs distinguish cold-start estimates, installed-runtime figures, offline demos, and host operation, and disclose scope; some figures still rely on author-provided records (2).

Risks and how to mitigate them
  • Running the plugins requires DeepSeek Harness; the offline demo covers the core path and does not establish host integration behavior.
  • The README shows v0.1.1, the CDN example pins v0.1.2, and package.json says 0.1.0; check the release artifacts and compatibility before installing.
  • Plugins may run with host privileges; review each organ's permissions and automatic recovery actions before deployment.
Evidence confidence: Low Reviewed Nov 08, 2026 Reviewed revision 9403444d87d5
See the full review method →
Maintain this project? Add the FARS badge to your README

The badge shows the current review score and verdict and links back to this page. It refreshes about once an hour after a re-review.

FARS

[![FARS review](https://followagents.com/badge/agent-body.svg)](https://followagents.com/en/agents/agent-body)
<a href="https://followagents.com/en/agents/agent-body"><img src="https://followagents.com/badge/agent-body.svg" alt="FARS review"></a>

Recent releases

Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.

  1. v0.1.1 v0.1.1 — portable message sources, reproducible packaging
    Release notes excerpt
    Portable message sources, and a release you can rebuild This release exists because of one failure mode: on harness 0.1.7-alpha.2 the organs stopped a turn with ``` 本轮运行失败 format v4 message requires a producer-owned source kind ``` What changed - Injected messages now carry the shape both session-format generations accept — { kind: 'plugin:<name>' } instead of the retired { kind: 'plugin', plugin: … } wrapper. The old wrapper is refused by format V4 on both the …
    View on GitHub ↗
  2. v0.1.0 v0.1.0 — 23 organs, one heartbeat
    Release notes excerpt
    Agent-Body v0.1.0 — 23 organs, one heartbeat The first public release of an organ-based plugin layer for [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness). Plugins here are not a tool list — they are organs in a living system, with senses, reflexes and a shared circulation. What's in the box 23 organs implementing one contract (sense → reflex → effect → homeostasis), across eight body systems: …
    View on GitHub ↗

FAQ

Do I need DeepSeek Harness just to try it?
No. npm run demo and npm run check work on a fresh clone without the host, network access, a model, an API key, or npm install. DeepSeek Harness is required for mounted organs and body_* runtime capabilities.
Does the 84.7% figure apply to the full prompt or API bill?
No. The 84.71% result covers only the cold-start block made from tool names, descriptions, and JSON Schemas. A development installation with real history measured 58% savings, and some capabilities require a second body_call.
Will it automatically retry every failed tool call?
No. It attributes the cause before applying a prescription. arg_error is explicitly excluded from automatic retry, side-effecting remedies wait for a higher-level decision, and wounds still open after five minutes become chronic.
Where does runtime data live?
Vitals, synapses, memory cards, and the pulse stream live in the Harness data home, not in the repository. The Harness GUI binds to 127.0.0.1, and the project warns against exposing it publicly or placing it behind a public reverse proxy.
Can every organ be installed with one command?
No. Ten self-contained organs have prebuilt tarballs. Other organs need host-side toolchains such as compilers, browsers, or external binaries and are installed through their replay scripts.
Is Agent-Body free?
Yes. Agent-Body is free to use and does not need a paid service.
Is Agent-Body open source?
Yes. The source code is on GitHub (1420079678-ctrl/agent-body) under the MIT licence.
Is Agent-Body safe to use?
FollowAgents' FARS review (Nov 08, 2026) scored Agent-Body 13/29 on trust and safety. None of the eight safety controls the review checks for was missing outright.
Is Agent-Body still maintained?
Development has slowed: the latest commit on GitHub is from 2026-09-30, within the last six months but not the last 30 days. The latest release is v0.1.1 (2026-09-23).
Who makes Agent-Body?
Agent-Body is published on GitHub by the account 1420079678-ctrl. FollowAgents has not verified which person or company is behind that account.
What is Agent-Body's official website?
The project's GitHub repository lists https://1420079678-ctrl.github.io/agent-body/ as its website. That link is set by the repository owner; the source code itself is on GitHub (1420079678-ctrl/agent-body).
Was this page helpful?
View on GitHub ↗ Install ↓

Related concepts