JSAT — AI Tool Integrations¶

JSAT works as an MCP server with every major AI coding tool. Each integration gives you the same depth of codebase intelligence: a launcher command, auto-connection on first use, a full set of MCP tools, and shell switch support.

For native Codex, native Claude, native OpenCode, direct Ollama, or a coding client launched through Ollama, start with the tabbed integration chooser. Those guides keep model ownership, local pulls, cloud sign-in, and managed lifecycle commands separate.

Feature Matrix¶

Feature	Claude Code	Codex	OpenCode	Cursor	Windsurf	Continue	Zed	Gemini CLI	Bob Shell
`jsat <tool>` launcher	✅	✅	via `jsat ollama --tool opencode`	✅	✅	—	✅	✅	✅
Auto-connect on launch	✅	✅	✅	✅	✅	—	✅	✅	✅
`jsat connect <tool>`	✅	✅	✅	✅	✅	✅	✅	✅	✅
`--scope project/global`	✅	Global only	Global only	✅	— (global)	— (global)	— (global)	— (global)	✅
Skills / custom commands	47 slash cmds	`$jsat` dispatcher	`/jsat` + `/jsat-help`	.cursorrules	.windsurfrules	10 custom cmds	.zed/JSAT.md	GEMINI.md	47 slash cmds + BOB.md
`switch <tool>` in shell	✅	✅	—	✅	✅	—	✅	✅	✅
`--keep-guidance` on disconnect	✅	✅	n/a	✅	✅	✅	✅	✅	✅
Tool type	CLI	CLI	CLI	GUI	GUI	IDE ext	GUI	CLI	CLI

Universal Behaviors (every connected tool)¶

Two behaviors ship with every /jsat <command> (or $jsat) dispatcher, regardless of which AI tool is connected — Claude Code, Codex, OpenCode, Cursor, Windsurf, Continue, Zed, Gemini CLI, and Bob Shell all get the same behavior because it's built into the generated dispatcher text, not into any one integration.

Input correction (on by default). Before routing, free-form ARGS are rewritten via jsat__prompt_rewrite to fix spelling, grammar, and phrasing without changing intent. Literal payloads — file paths, diffs, code blocks, git refs, URLs — are left untouched so a misspelled path like fix src/paymnet/service.py still resolves correctly. If the rewrite materially changes ARGS, the dispatcher tells you what changed before routing. Pass raw=true on any call to skip this step and use ARGS exactly as typed.

Learning module (after every command). Once a command's reply is given, the dispatcher asks whether anything durable was learned — most calls yield nothing new and are skipped silently. A concrete, project-specific fact (a gotcha, a verified false positive, a non-obvious dependency) is saved to the project's knowledge base via jsat__knowledge_add. A concrete gap in JSAT's own tools or skills is saved instead to the jsat improve backlog, feeding future jsat improve runs. Either way you'll see a one-line Learned: ... note naming where it was saved.

Quick Start (any tool)¶

# 1. Index your codebase
jsat index .

# 2. Open your AI tool with JSAT pre-loaded
jsat claude      # Claude Code
jsat codex       # OpenAI Codex CLI
jsat opencode    # OpenCode (native, project scope)
jsat ollama --tool opencode  # OpenCode through Ollama; auto-connects JSAT
jsat cursor      # Cursor IDE
jsat windsurf    # Windsurf
jsat gemini      # Gemini CLI
jsat zed         # Zed editor
jsat bob         # Bob Shell

Each launcher auto-connects JSAT if not already wired and opens the tool with 69 MCP tools ready.

Managed lifecycle¶

jsat start                           # Claude + Codex + OpenCode, one terminal each
jsat start opencode --via ollama     # one client, Ollama selector
jsat start --model <model>-cloud   # all three through one Ollama model
jsat ps
jsat restart                         # Claude + Codex + OpenCode
jsat stop                            # every running managed client
jsat resume                          # Claude + Codex + OpenCode
jsat resume codex --session ID       # one named Codex session

all is the default target. Since interactive TUIs cannot share one terminal, multi-client start, restart, and resume open a separate terminal window for each client. Supported terminal launchers are x-terminal-emulator, GNOME Terminal, Konsole, and Xfce Terminal. An explicit target runs in the current terminal.

JSAT stores a minimal process record under ~/.jsat/runtime/. stop validates the recorded process-start identity before sending a signal, then stops that process and its current descendants. It never kills processes merely because their executable is named claude, codex, or opencode. Use --force only when graceful shutdown times out.

jsat resume resumes AI-client conversation history. jsat session resume remains the separate command for interrupted JSAT skills such as magic and crack.

Per-Tool Details¶

OpenCode through Ollama¶

jsat index .
jsat connect ollama tool=opencode           # configure only OpenCode
jsat ollama --tool opencode                  # choose local or cloud interactively
jsat ollama --tool opencode -m qwen2.5:0.5b    # local
ollama signin
jsat ollama --tool opencode -m <model>-cloud     # cloud

The launcher writes JSAT's MCP entry to the repo's .opencode/opencode.json (pinned to that repo) before running ollama launch opencode. It also installs /jsat and /jsat-help under .opencode/commands/. Ollama's inline model configuration is deep-merged with that file. A standalone OpenCode installation is not required. For configuration without launch, run jsat connect opencode, then run bare ollama and choose OpenCode → sign in if prompted → choose a model. Pass --global to jsat connect opencode to target ~/.config/opencode/ instead of a project.

jsat connect ollama configures every Ollama-launched client JSAT currently supports: Claude, Codex, and OpenCode. Use --tool opencode, opencode, or tool=opencode to configure only one. JSAT does not configure unknown menu entries because each client has a different MCP configuration format.

Provider routing remains separate for every mode:

Native Claude, Codex, and OpenCode use claude_cli, codex_cli, and opencode_cli, respectively.
Ollama-launched OpenCode reads OPENCODE_CONFIG_CONTENT; Ollama-launched Claude reads the ANTHROPIC_DEFAULT_*_MODEL launch variables.
The selected Ollama model can be local or cloud. It overrides .jsat/config.yaml only for that launched process and does not require another pull.

Claude Code¶

jsat claude                              # open with JSAT tools
jsat connect claude --global             # global — all sessions (recommended)
jsat connect claude                      # project scope only
jsat connect claude --no-skills          # MCP only, skip slash commands

What gets installed (global): - ~/.claude/settings.json — MCP server config - ~/.claude/commands/jsat-*.md — 47 slash commands

What gets installed (project): - .claude/settings.json — MCP server config - .claude/commands/jsat-*.md — 47 slash commands

Slash commands: /jsat-query, /jsat-blast-radius, /jsat-security, /jsat-review, /jsat-test-gaps, /jsat-knowledge, /jsat-incident, /jsat-prompt, /jsat-tokens, /jsat-ithinking, and 21 more.

In the JSAT shell:

switch claude-cli    → launch full Claude Code session

OpenAI Codex CLI¶

jsat codex                               # open with JSAT pre-loaded
jsat connect codex                       # global MCP config + $jsat skill
jsat codex --repo /path/to/repo           # launch Codex in a specific repo
jsat codex resume <session-id>            # resume an existing Codex session

What gets installed: - ~/.codex/config.toml — one [mcp_servers.jsat] entry - ~/.codex/skills/jsat/SKILL.md — one Codex skill dispatcher for $jsat

JSAT does not generate .codex/, AGENTS.md, or .agents/skills inside the target repo for Codex. Run Codex from the repo you want to analyze, or use jsat codex --repo /path/to/repo; JSAT resolves that working directory at runtime. Any extra arguments after jsat codex are forwarded to the real Codex CLI.

Use in Codex: run $jsat magic TASK, $jsat query QUESTION, or ask naturally, e.g. "use JSAT to trace the blast radius of src/payment/service.py". The dispatcher also treats @jsat magic TASK as the same JSAT request when Codex routes it to the skill. Direct MCP tools such as jsat__query, jsat__blast_radius, jsat__security_review, jsat__get_test_gaps, and jsat__submit_for_review remain available.

In the JSAT shell:

switch codex    → launch Codex CLI session

Install Codex: curl -fsSL https://chatgpt.com/codex/install.sh | sh

OpenCode¶

jsat opencode                             # open with JSAT pre-loaded (project scope)
jsat opencode --repo /path/to/repo         # launch OpenCode in a specific repo
jsat opencode resume <session-id>          # resume an existing OpenCode session
jsat connect opencode --global             # machine-wide MCP instead of per-repo

What gets installed (project scope): - .opencode/opencode.json — MCP server config, pinned to the repo via --repo - .opencode/commands/jsat.md + jsat-help.md — /jsat slash commands - AGENTS.md — marked JSAT guidance block (opencode reads it automatically)

jsat opencode auto-connects JSAT at project scope on first use, then runs the real OpenCode binary from the repo directory so it picks up .opencode/. Any extra arguments after jsat opencode are forwarded to the real OpenCode CLI.

In the JSAT shell:

switch opencode    → launch OpenCode session

Install OpenCode: npm install -g opencode-ai — or use jsat ollama --tool opencode when Ollama manages the install.

Cursor¶

jsat cursor                              # open Cursor with JSAT pre-loaded
jsat connect cursor                      # global scope (default)
jsat connect cursor --scope project      # project scope (.cursor/mcp.json)

What gets installed: - ~/.cursor/mcp.json (global) or .cursor/mcp.json (project) — MCP server - .cursorrules — JSAT tool guidance (Cursor reads from project root)

In the JSAT shell:

switch cursor    → open Cursor in background

Install Cursor: macOS: brew install --cask cursor · Linux/Windows: cursor.com/download

Windsurf¶

jsat windsurf                            # open Windsurf with JSAT pre-loaded
jsat connect windsurf                    # wire JSAT in

What gets installed: - ~/.codeium/windsurf/mcp_config.json — MCP server - .windsurfrules — JSAT tool guidance (Windsurf reads from project root)

In the JSAT shell:

switch windsurf    → open Windsurf in background

Install Windsurf: macOS: brew install --cask windsurf · Linux/Windows: windsurf.ai/download

Continue.dev¶

Continue is an IDE extension (VS Code, JetBrains). There's no jsat continue launcher since it runs inside your IDE.

jsat connect continue                    # wire JSAT in + install 10 /jsat-* commands

What gets installed: - ~/.continue/config.json — MCP server + 10 customCommands

Custom commands (curated subset): /jsat-query, /jsat-blast-radius, /jsat-security, /jsat-review, /jsat-test-gaps, /jsat-knowledge, /jsat-incident, /jsat-prompt-rewrite, /jsat-tokens, /jsat-ithinking.

After connecting, reload Continue: Cmd/Ctrl+Shift+P → Continue: Reload.

Zed¶

jsat zed                                 # open Zed with JSAT pre-loaded
jsat connect zed                         # wire JSAT in

What gets installed: - ~/.config/zed/settings.json — context server config - .zed/JSAT.md — JSAT tool guidance as project context

In the JSAT shell:

switch zed    → open Zed in background

Install Zed: macOS: brew install --cask zed · Linux: curl -f https://zed.dev/install.sh | sh · Windows: not yet available

Google Gemini CLI¶

jsat gemini                              # open Gemini CLI with JSAT pre-loaded
jsat connect gemini                      # wire JSAT in

What gets installed: - ~/.gemini/settings.json — MCP server config - GEMINI.md — JSAT tool guidance (Gemini CLI reads from project root automatically)

In the JSAT shell:

switch gemini    → launch Gemini CLI session

Install Gemini CLI (all platforms): npm install -g @google/gemini-cli

Bob Shell¶

See Bob Shell for the full standalone guide.

jsat bob                                 # open Bob Shell with JSAT pre-loaded
jsat bob --mode advanced                 # open in specific mode (plan, code, advanced, ask)
jsat connect bob --global                # global — all sessions (recommended)
jsat connect bob                         # project scope only
jsat connect bob --no-commands           # MCP + BOB.md only, skip slash commands

jsat bob opens a clean interactive Bob session; JSAT tools and guidance are loaded from .bob/ and BOB.md, so nothing is injected as a throwaway prompt.

What gets installed: - .bob/settings.json (project) or ~/.bob/settings.json (global) — MCP server config - .bob/commands/jsat-*.md (or ~/.bob/commands/) — 47 /jsat-* slash commands - BOB.md — JSAT tool guidance (Bob Shell reads from project root)

Slash commands: type / in Bob Shell to browse them — /jsat-query, /jsat-blast-radius, /jsat-security, /jsat-review, /jsat-test-gaps, /jsat-prompt, /jsat-ithinking, and 24 more. /jsat-prompt optimizes your query and then answers it (use --optimize-only to just see the rewrite).

In the JSAT shell:

switch bob    → launch Bob Shell session

Install Bob Shell (all platforms): npm install -g @ibm/bob-shell

Bob Shell modes: - plan — Planning and design mode - code — Code implementation mode - advanced — Advanced code mode with more tools - ask — Question and answer mode

JSAT Shell — `switch` Reference¶

From inside jsat shell, you can switch to any tool:

switch claude-cli   → full Claude Code + JSAT MCP (recommended for Claude)
switch codex        → Codex CLI (reads ~/.codex/config.toml)
switch gemini       → Gemini CLI (reads ~/.gemini/settings.json + GEMINI.md)
switch cursor       → open Cursor IDE in background
switch windsurf     → open Windsurf IDE in background
switch zed          → open Zed in background
switch bob          → Bob Shell session
switch gpt <model>  → OpenAI API in JSAT shell (needs OPENAI_API_KEY)
switch ollama <model>    → Ollama in JSAT shell
switch anthropic <model> → Claude API in JSAT shell (needs ANTHROPIC_API_KEY)

Disconnect¶

jsat disconnect claude                   # Claude project scope
jsat disconnect claude --scope all       # Claude everywhere
jsat disconnect codex                    # Codex
jsat disconnect opencode                 # OpenCode project scope (.opencode/)
jsat disconnect cursor                   # Cursor (global + project)
jsat disconnect windsurf                 # Windsurf
jsat disconnect continue                 # Continue (removes commands too)
jsat disconnect zed                      # Zed
jsat disconnect gemini                   # Gemini
jsat disconnect bob                      # Bob Shell
jsat disconnect all                      # every tool at once

# Keep guidance files for integrations that generate them (.cursorrules, etc.)
jsat disconnect cursor --keep-guidance

List Active Connections¶

jsat connect list       # show every tool that has JSAT wired
jsat doctor             # full health check including connected tools

MCP Tools Available in Every Tool¶

All 69 JSAT MCP tools are available to every connected AI tool:

Category	Key tools
Graph exploration	`jsat__query`, `jsat__get_function`, `jsat__get_class`, `jsat__trace_call_chain`
Impact analysis	`jsat__blast_radius`, `jsat__blast_radius_diff`, `jsat__blast_radius_file`
Security	`jsat__security_review`, `jsat__list_secrets`, `jsat__get_auth_coverage`
Code quality	`jsat__submit_for_review`, `jsat__get_test_gaps`, `jsat__generate_unit_test`
Knowledge	`jsat__knowledge_query`, `jsat__knowledge_add`, `jsat__generate_runbook`
Investigation	`jsat__investigate_incident`, `jsat__get_recent_changes`
Prompt tools	`jsat__prompt_optimize`, `jsat__prompt_multi_agent`, `jsat__prompt_diff`
Token tools	`jsat__token_count`, `jsat__token_compress`, `jsat__token_budget`
Migration	`jsat__validate_migration`, `jsat__suggest_zero_downtime`
IThinking	`jsat__ithinking_plan`, `jsat__ithinking_reflect`

JSAT — AI Tool Integrations¶

Feature Matrix¶

Universal Behaviors (every connected tool)¶

Quick Start (any tool)¶

Managed lifecycle¶

Per-Tool Details¶

OpenCode through Ollama¶

Claude Code¶

OpenAI Codex CLI¶

OpenCode¶

Cursor¶

Windsurf¶

Continue.dev¶

Zed¶

Google Gemini CLI¶

Bob Shell¶

JSAT Shell — switch Reference¶

Disconnect¶

List Active Connections¶

MCP Tools Available in Every Tool¶

JSAT Shell — `switch` Reference¶