Skip to content

Latest commit

 

History

95 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Agent Doctor by NexoToken

简体中文 · Install · NexoToken setup guides · Compatibility · Report an issue

CI License Public beta

Agent Doctor local evidence flow

A local reliability and guidance layer for AI coding agents.

Agent Doctor is an open-source local tool by NexoToken, an AI API and agent tools platform for developers.

Codex and Claude Code do the work. Agent Doctor watches the evidence, detects when a task is looping, losing context, or trying to finish without validation, and sends a bounded next-step instruction back to the running agent. It also explains why a task became slow, expensive, repetitive, or unreliable. Missing data stays unavailable instead of becoming a fabricated zero.

Its shared project brain also carries work across clients. When you leave a captured Codex task and open Claude Code in the same repository, Agent Doctor automatically injects a bounded handoff containing the latest goal and result, confirmed project memory, source provenance, and explicit limitations. It does not paste the full transcript, and every delivery is visible in the local dashboard.

The guidance engine is deterministic, runs locally, and does not call another model. Raw prompts, source files, commands, tool inputs, and tool results are not used by the guidance rules; supported hooks retain only bounded labels and non-reversible evidence fingerprints.

60-second local start

Clone the repository and run one command. It installs dependencies, builds and tests the product, installs the binary, configures the owned Codex and Claude Code assets, opens the dashboard, and keeps the local service running:

./scripts/install-local.sh

start checks and idempotently prepares the Agent Doctor-owned Codex integration, refreshes detected clients, starts the loopback services, and opens the dashboard in the default browser. Use start --no-open when you want to copy the printed URL yourself. Agent Doctor never binds a public interface. The dashboard includes Overview, Task evidence, Costs, Memory, Comparison, Trends, Integrations, and Privacy.

The simplest live capture path reuses the API base URL already configured in your terminal and injects the local proxy only into the child process:

agent-doctor start                              # check integrations and open dashboard
agent-doctor run -- codex                       # start Codex with live capture
# or: agent-doctor run -- claude

AGENT_DOCTOR_UPSTREAM_URL has priority when a client does not expose its configured base URL through OPENAI_BASE_URL or ANTHROPIC_BASE_URL. A client that was already running must be restarted through the wrapper; Agent Doctor does not attach to arbitrary existing processes or rewrite credentials.

The published v0.1.0-beta.1 archives and checksums passed independent verification. Install the public beta with:

curl --fail --proto '=https' --tlsv1.2 https://raw.githubusercontent.com/18534516725/Agent-Doctor/main/install.sh | sh
irm https://raw.githubusercontent.com/18534516725/Agent-Doctor/main/install.ps1 | iex

Both installers download only the official GitHub release, verify its SHA-256, install the binary, and run the consented setup step. See installation and uninstallation before using a remote installer.

First diagnosis

agent-doctor doctor --json
agent-doctor diagnose --json
agent-doctor costs --json
agent-doctor dashboard --no-open

The first dashboard panel is the Task Guardian: it shows whether work is on track, needs redirection, is waiting for validation, or has been blocked by a capable hook. The result states evidence provenance and precision. A diagnosis is not a verdict: unsupported evidence and absent billing records remain visible limitations.

Runtime guidance boundary

  • Claude Code: official hooks can return guidance and, in guard or autopilot, enforce supported PreToolUse or unverified Stop decisions.
  • Codex: MCP and Skill assets provide evidence-backed guidance at task checkpoints. The client can ignore MCP/Skill text, so this is not a deterministic block.
  • Other clients: capability depends on the public interface listed in the compatibility matrix. Agent Doctor never claims enforcement where only observation or MCP advice exists.

Project control levels are observe (record only), guide (advise), guard (enforce supported high-confidence rules), and autopilot (strongest supported local controls). The default is guide.

Public commands

Command Purpose
agent-doctor setup --json Preview detected clients and exact owned changes
agent-doctor setup --yes --json Apply the reviewed Codex MCP block
agent-doctor setup --all --yes --json Install owned Codex MCP/Skill/AGENTS and Claude Hook/Skill assets
agent-doctor start Check managed integrations, start loopback services, and open the dashboard
agent-doctor start --no-open Same startup flow without opening a browser
agent-doctor dashboard Alias for the visual workspace
agent-doctor diagnose --json Summarize available task evidence
agent-doctor compare --json Report matched-cohort readiness/results
agent-doctor context --json Report bounded memory state without content
agent-doctor costs --json Separate exact, estimated, and unavailable cost
agent-doctor doctor --json Check database and detected clients
agent-doctor pause --json Persistently pause local lifecycle capture
agent-doctor pause --resume --json Resume local lifecycle capture
agent-doctor export --json Export sanitized aggregates, never event payloads
agent-doctor forget --yes --json Delete the local Agent Doctor database
agent-doctor run -- <command> Start a client through the local capture proxy without shell evaluation
agent-doctor uninstall --yes --json Remove only Agent Doctor-owned Codex config
agent-doctor version Print the installed version

Supported clients

Codex, Claude Code, Cline, OpenCode, Cursor, Windsurf, Roo Code, Continue, Aider, Cherry Studio, and generic command-line tools have declared capability contracts. “Supported” does not mean that a client exposes every signal. See the compatibility matrix for exact A/B/C capability levels and installation boundaries.

Cost truthfulness

  • exact: a charge reported by a compatible billing source;
  • estimated: local token usage multiplied by a versioned public catalog;
  • unavailable: the required evidence does not exist.

These values are never merged into a misleading total. Currency conversion requires a versioned rate. Read the cost methodology.

Local privacy

The SQLite database lives in the current user's configuration directory with user-only permissions. When live capture is enabled, complete user, assistant, system, and tool messages are stored locally so the owner can inspect the real conversation. API keys, Authorization headers, cookies, and transport headers are forwarded in memory and never written to SQLite. Replay is disabled until the user approves the exact hashed plan, base commit, commands, call limit, and cost limit. Read the privacy model.

Architecture

documented client interface → sanitizer → local event contract → SQLite
           ↑                                         ↓
   Hook / MCP / Skill ← deterministic guidance ← evidence fingerprints
                                                     ↓
                         safe aggregates → loopback Task Guardian

Limitations

  • setup --all owns only the documented Agent Doctor files/marked blocks for Codex and Claude Code, creates backups, and is safe to run repeatedly.
  • Descriptive comparison starts with two captured sessions. A cohort-level result still requires at least 15 matched samples and never invents a winner.
  • Agent Doctor does not read private client databases or bypass client permissions.
  • No public release or ranking claim is valid until the tagged GitHub workflow completes and the published checksums are independently verified.

Documentation

License

Apache License 2.0. See LICENSE.

Public beta feedback

Real projects are more useful than synthetic demos. If a diagnosis, cost state, client boundary, or handoff is wrong, choose the matching issue form. Never attach a complete SQLite database, full transcript, credential, private source file, or authorization header. See the feedback guide.

About

Local-first observability, cost analysis and project memory for Codex, Claude Code and AI coding agents.

Topics

Resources

Contributing

Security policy

Stars

2 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages