base-context — made in 
by Synerise AI.
Keep the work. Focus the context.
Give your agent a workspace, not just a conversation. Base Context combines persistent Python, retained outputs, and instruction recovery to keep coding and research moving across long sessions. Built by Synerise, on Prime Agent.
27–39% shorter mean attempt times. 150/150 native completions. In our curated 50-task benchmark, Base was also faster in 128 of 141 pairs where both agents completed and fully passed. Explore the results.
Install · Workflow · Benchmarks · Design · Documentation · Attribution
On macOS or Linux:
curl -fsSL https://github.com/BaseModelAI/base-context/releases/latest/download/install.sh | bashThe installer offers missing prerequisites and prepares Node.js, uv, and managed Python. Run the final PATH command it prints, then:
cd /path/to/your/project
base-contextBring your existing provider account. Choose /login, then /model and /effort. Have ChatGPT with Codex access? Choose OpenAI Codex in /login—no separate API key needed. Other supported subscriptions and API keys are available through their provider routes. Provider setup.
This page describes 1.1.3; install commands use the latest public release. What's new.
Prefer npm, Windows, or a source build?
For an existing Node.js/npm installation, use Node.js 22.12+ on 22.x, or 23.3+. In Bash/Zsh:
npm install -g @ponythewhite/base-context
BASE_CONTEXT_INSTALL_UV=1 base-contextThe first normal launch prepares managed Python; the flag permits installing missing uv. Use this route instead of the installer above.
Windows/PowerShell, source builds, updates, rollback, and uninstall.
The persistent Python workspace keeps parsed data, variables, and command handles ready for the next step. Read a dataset, inspect a repository, start a command—then work with the result instead of recreating it. The agent runs project tools in the project's own environment. You do not need to write Python yourself.
A long log belongs in retained history, not in every model request. Base Context keeps the result and lets the agent recover a specific failure, passage, or line with prime_context. The useful evidence stays within reach while the conversation moves forward.
A TaskFrame carries selected earlier user instructions and goal state alongside summaries, without repeating user text already visible in the request. Retained sources support recovery after a saved-session restart, too. Compact the conversation and keep working from the original details when they matter.
Give the API review to one worker and the documentation to another. Each gets its own context and reports back through messages or files. The parent can keep working on the main task while those results arrive.
These capabilities are available in ordinary sessions. Start with the ten-minute workflow.
Coming from Codex? Start in your repository, keep project rules in AGENTS.md, and ask for an outcome:
Fix the failing parser test. Keep the public API unchanged and explain the change.
For a longer job, give the agent a persistent goal:
/goal Implement the migration in PLAN.md and run the project checks
An active goal keeps work moving across turns. Use /goal status, /goal pause, /goal resume, or /goal clear to manage it. Goals and background work.
To split the work:
Delegate the API review and documentation update to separate workers.
Keep working on the parser fix, then integrate their replies.
/agents 4 sets the subagent cap; /agents shows it. The default is four. Open base-context agents to see your workers.
Want to ask a question without steering the main task? Use /btw Why did you choose this parser? for a separate, tool-free side conversation. Esc returns to the main editor.
| Want to… | Use |
|---|---|
| Add a file or run a command | @path, !command; !!command keeps output out of model context |
| Steer running work / queue a later request | Enter / Alt+Enter |
| Inspect context, tokens, and cost estimates | /context or /usage |
| Summarize context / refine saved harness advice | /compact / /refine |
| Continue the latest saved session | base-context -c |
| Reattach to a resident agent | base-context attach <agent> |
| Stop one agent / all agents and services | base-context stop <agent> / base-context shutdown |
Normal interactive sessions can keep running after the terminal detaches. Reattach when you're ready. Full usage.
Our curated 50-task comparison puts Base Context development builds and Codex 0.160.0 through staged work with forced compaction and scheduled cold restarts. Across three matching model/effort profiles, Base delivered 27–39% shorter mean attempt times.
| Profile · 50 attempts per harness | Mean time, Base / Codex | Native completions, Base / Codex | Full artifact passes, Base / Codex |
|---|---|---|---|
| GPT-6 Astra · medium | 508.4 / 697.2 s | 50 / 50 | 49 / 50 |
| GPT-6.1 Sol · high | 606.4 / 967.4 s | 50 / 50 | 50 / 49 |
| GPT-6.1 Sol · xhigh | 908.1 / 1,481.3 s | 50 / 44 | 49 / 47 |
The speed difference also holds among successful runs: Base was faster in 128/141 pairs where both harnesses completed natively and fully passed. Across all selected attempts, Base completed 150/150 versus 144/150, with 148/150 full artifact passes versus 146/150. Artifact scores include three passing Codex timeouts; completion is scored separately.
About 11.3% lower estimated cost overall: $1.675 for Base versus $1.888 for Codex per task, a saving of $0.213 per task across the 150 selected attempts per product.
| Model / effort | Base estimate / task | Codex estimate / task | Base saving / task | Base saving |
|---|---|---|---|---|
| Astra / medium | $3.325 | $3.668 | $0.343 | 9.3% |
| Sol 6.1 / high | $0.753 | $0.836 | $0.082 | 9.8% |
| Sol 6.1 / xhigh | $0.948 | $1.161 | $0.213 | 18.3% |
| All profiles | $1.675 | $1.888 | $0.213 | 11.3% |
Both sides include missing-cost estimates: comparable-request means for Base's missing usage, and requested-profile model rates for Codex's missing model identities. These are API-list-rate estimates for captured requests, not subscription bills or complete family spend. Supporting components, assumptions and sensitivity.
Scope: 50 curated and outcome-informed tasks, three profiles, first provider-clean attempt per cell. Clean failures and timeouts are included. Request-budget selection was off. Full methodology covers task selection, provider exclusions, mixed builds, and cost sensitivity. Download the benchmark summary.
Earlier comparisons remain withdrawn and available as archives: Python 30 · Earlier evaluation records.
Keep a notebook. Work from a clear desk. Retained history is the notebook; the model's working context is the desk. Base Context separates them so the agent can keep its work without carrying every byte into every request.
| Layer | What it brings to the workflow |
|---|---|
| Python workspace | Reusable variables, parsed data, skills, and command handles |
Retained history and prime_context |
Original public records, ready for targeted recovery |
| TaskFrame | Selected earlier user instructions and goal state alongside summaries |
| Optional request-budget profile | Dependency-aware selection and stable context choices on supported routes |
| TypeScript host | Model calls, sessions, goals, scheduling, provider recovery, and child lifecycles |
Workers start with await rlm(...) and return results through messages or files. Recognized transient provider errors can recover within the same invocation without replaying completed tools.
Want tighter request control? Enable request-budget selection with an explicit enforce-mode route/model profile in settings or the SDK. It is opt-in and applies per request, separate from goal budgets. Working configuration.
Built on Prime Agent, focused on context. Prime Agent gives us the persistent Python REPL, recursive delegation, skills, messaging, goals, and long-running sessions. Base Context develops framed history, indexed recovery, TaskFrames, dependency-aware selection, and stable context epochs around that foundation. Why we forked · Context design.
Save important deliverables in files; Python restoration is best-effort. Commands run with your user permissions, so use an external sandbox for untrusted code. Security guidance.
Quickstart · Usage · Settings · Providers · Skills · RLM · SDK · All docs · 1.1.3 release notes
See the changelog for releases. Contributions are welcome: CONTRIBUTING.md.
Synerise develops Base Context. BaseModelAI/base-context is its GitHub home; @ponythewhite/base-context is its npm package; base-context is the command.
MIT, with upstream notices preserved in NOTICE.
Prime Agent is an excellent foundation, and we are grateful to its authors. Thank you to Prime Intellect for the Python-first agent and recursive runtime, and to Mario Zechner's Pi, on which Prime Agent builds. Base Context is an independent Synerise fork.