Describe a task in plain language and 13x builds a budgeted, local-only context packet from your code map — then pastes it straight into your running agents so they read the right files first.
Context Injection lets you describe a task in plain language and have 13x assemble a focused context packet — the most relevant files, symbols, dependency neighbors, ranking reasons and a suggested inspection order — then paste that Markdown directly into one or more running agent terminals. Packets are built entirely from your on-disk code map: no code upload, no external AI call. Everything is saved under .13x/context-packets/, and each target slot gets a compact PKT badge in its terminal header that opens full packet details on click.
build_auto_context, generate_context_packet and send_context_packet_to_slots are Tauri commands that operate only on your on-disk code map and PTYs. No code is uploaded, no external AI API is called, and no hidden commands are sent.
Four budgets — TINY, SMALL, BALANCED, DEEP — cap files, symbols and neighbors (balanced is 8 files / 30 symbols / 6 neighbors). Each packet shows estimated full-repo tokens vs packet tokens and a reduction percent so you see what you save.
One panel composes multiple task rows, each targeting different agent slots with a Config Advisor recommendation. On SEND ALL, claude slots get their chosen config injected as slash pre-commands while foreign CLIs receive the bare packet — a config never mutates the wrong agent.
Agents receive plain, human-readable Markdown, and the selector skips sensitive paths like .env, .env.* and secrets.* so credentials are never bundled into a packet. Only the DEEP budget inlines source, and even then it's bounded and still skips sensitive files.
After a send, each slot shows a PKT badge with file/symbol counts that survives restarts via local usage logging and rehydration. It flips to PKT* when the code map is rescanned after the packet was built, and shows CTX ERR on a failed send.
If you previewed or created a packet for exactly the task you're sending, that same packet is reused instead of regenerating a near-duplicate — what you saw in the preview is what the agent gets.
Describe the task
Open the AUTO CONTEXT panel (TASK in the top bar, or CONTEXT on an agent card) and type a task in plain language, like 'implement group resume for all terminals'. Pick a budget and toggle whether to include neighbor files, symbols and constraints.
Preview the selection
PREVIEW CONTEXT ranks your code map for the task and shows the top files, symbols, dependency neighbors, ranking reasons, a suggested inspection order, and token + reduction estimates. Flip between a structured summary and the raw packet Markdown.
Create the packet
CREATE PACKET writes a human-readable .md and a .json metadata file under .13x/context-packets/. Copy it to the clipboard, or reload any of your recent packets from the Code Map PACKETS list.
Send to your agents
SEND TO AGENT or SEND ALL pastes the packet into every selected running slot, runs any per-slot config pre-commands, logs usage and sets the PKT badge. Leave 'send enter after paste' off to paste without auto-submitting, or on to fire immediately.
Inspect any time
Click a slot's PKT / PKT* / CTX ERR badge to see the task, packet id, budget, file/symbol/neighbor counts, token and reduction estimates, send time, status, and the exact on-disk packet and metadata paths — then copy the packet back out.
No. Packets are built entirely from the local on-disk code map. No code is uploaded, no external AI API is called, and no hidden commands are sent. Packets, metadata and the usage log all live under the project's .13x/context-packets/ directory.
Yes — you need a scanned 13x Code Map. With no map, preview and send fail with 'code map not found. scan map first.' and the panel surfaces an inline SCAN MAP button to fix it in one click.
They're rough estimates derived from character counts (roughly chars per token), not a real tokenizer. They're there to show relative reduction, not to be exact billing figures.
Not in V1 — packets reference files, symbols, neighbors and an inspection order rather than pasting whole files. Only the DEEP budget inlines source excerpts, and those are bounded (about 8 files, 120 lines each) and still skip sensitive paths.
13x is local-first and free of telemetry — your code, commands and shells never leave the machine. Join the waitlist for the next build.