Local-first AI terminal · Claude · Codex · Gemini

Termpolis

Stop re-explaining your codebase to AI.

Claude, Codex and Gemini share one memory in Termpolis, and it learns as you work. A new terminal opens already knowing your project, your decisions and what got figured out yesterday. The memory is stored on your own computer.

Windows · macOS · Linux  ·  Free and open source (Apache 2.0)  ·  No account  ·  iPhone

3 agents, 1 memory What Claude learns, Codex and Gemini can recall
0.97 recall@10 The right memory is in the top 10 for 97% of test queries
~20 ms lookup Local search, nothing leaves your machine
Tested in CI The build fails if recall or speed slips. Reproduce it

🧠 Mneme · the memory that learns

One memory for all your agents

Every AI session normally starts cold, so you re-explain the project and pay tokens to reload it. Mneme reads your past sessions, keeps what matters, and gives every agent the relevant part when it needs it.

  • Learns from every session. When a Claude or Codex agent finishes a task, the fix, the decision or the gotcha is saved as a lesson the others can use. Gemini is asked to record its own.
  • Gets better with use. Memories that help rank higher, ones that don't fade, and contradictions between agents are flagged.
  • Knows your code. A local code graph lets agents find callers and see what a change would break before they edit.
  • Private by default. Embeddings run on your computer, the store is encrypted with a key in your OS keychain, and you can export it all to a .zip.

How the memory works →

How Mneme works: Claude, Codex and Gemini write to and recall from one local memory that distills lessons from each session.

⚡ Token Headroom · spend less per task

Get more work out of every session

The biggest cost in a long session isn't your prompts. It's the file reads, command output and screenshots sent back to the model on every turn. Termpolis cuts that on your computer before it leaves.

  • Compression on by default for Claude Code. A local proxy shrinks tool output and pasted images, never your words or the model's replies, and keeps the prompt cache intact.
  • The right model for each message. Turn it on and each message goes to the model and effort it needs, decided on your machine in milliseconds.
  • A fresh session when it pays. When starting over would halve the cost of each turn, the pane offers a Fresh session that carries a handoff note forward.
  • An honest receipt. Savings are shown against compressible text, all input and your whole bill, so a big percentage is never mistaken for a big discount.

How Token Headroom works →   Model per message →

🛡 AI Security Center

Safer use of hosted models

Any tool that talks to a hosted model sends your prompt to that provider, and Termpolis can't change that. What it can do is catch secrets where they actually escape and check what you install. Every check runs locally and every log stays on your computer.

  • Commit Shield. A commit or push that carries a key is blocked, not just logged: from the Git panel out of the box, and from any terminal or IDE once you install its git hooks.
  • Secrets never enter the memory. They are scrubbed before anything is embedded or written.
  • Safe Import. Third-party skills, plugins and MCP servers are scanned before they install. A red result can't install.
  • Prompts watched. What you type to an agent is checked against 97 secret rules. A hit is recorded and named so you know what to rotate. It doesn't block the prompt, and we say so.

What each check does and doesn't do →

Where Termpolis checks for secrets: Commit Shield blocks a key in a commit or push, the memory scrub removes secrets before anything is stored, Safe Import gates add-ons before they install, and the prompt watch detects and names a secret in a prompt without blocking it.

📱 Termpolis Remote · 🔗 Linked machines

Your agents, wherever you are

Both are off until you turn them on, and both are encrypted end to end. The relay passes sealed messages and can't read your terminals, prompts or answers.

Termpolis Remote for iPhone

See the terminals running on your computer and answer an agent's question from your phone, so a blocked task becomes a ten-second reply. The desktop decides what the phone may do, nothing applies until you confirm the pairing's safety words on the desktop, and nothing runs on the phone.

Free to install, iOS 16.4+. Relay access $4.99 a month, first week free. Android follows.

App Store Setup & support

Linked machines

Link two Termpolis computers and an agent on one can have Claude, Codex or Gemini on the other do a task and hand back the answer. Ask Claude on your laptop to “have Codex on linux run the integration tests and summarise the failures.”

Read-only by default. Editing files and running commands stay off until you allow them, per computer.

How Linked machines works

And a full terminal around it

Everything else you get

Your real Claude Code, Codex and Gemini, unchanged, plus the tools for running several at once.

Second Opinion

Have a different agent and model check the last answer. The reviewer is read-only.

Docs →

Cross-AI handoff

Move a Claude session to Codex or Gemini without re-explaining it.

Docs →

Multi-agent swarm

Describe a goal and a local conductor splits it across up to 8 agents, then you review the result.

Docs →

Workflow Orchestrator

Chain agents, commands and skills into one repeatable run, with branches and loops.

Docs →

See what agents are doing

An activity feed, context pins, a redundancy detector and per-agent efficiency stats.

Docs →

Git panel and diffs

Stage, read diffs, commit, pull and push, and open any diff an agent prints.

Docs →

40 MCP tools

Agents can drive terminals, coordinate a swarm, and search the memory and code graph.

Docs →

Share-ready output

Copy terminal output cleanly for Teams, Slack or a pull request with one shortcut.

Docs →

Terminal basics, done well

Split panes, saved workspaces, themes, keybindings and autocomplete, on PowerShell, cmd, Git Bash, bash or zsh. On Windows your keystrokes go straight to the shell, so typing stays quick while the app is busy.

Docs →

⚖ How Termpolis compares

Termpolis vs other AI terminals

Every competitor cell is sourced from that tool's own public documentation as of October 2026. Marks: ✅ ships, ⚠ partial / opt-in / third-party, ❌ not available.

Capability Termpolis Warp Wave Terminal Air app (JetBrains) Tabby
Open source & self-hostable
Source available, you can audit/fork
✅ Apache 2.0 ✅ AGPL v3 client (Apr 2026) ✅ Apache 2.0 ❌ Closed source (JetBrains) ✅ MIT
No required cloud account
Works the day you install it, no signup
✅ No Termpolis account ⚠ No login to install; AI + sync need an account ✅ No account; free AI needs telemetry ⚠ JetBrains Account only for JetBrains AI, Junie & cloud ✅ No account
Multi-agent swarm (parallel agents w/ live conductor)
Runs Claude + Codex + Gemini together
✅ Local conductor runs Claude, Codex and Gemini together ⚠ Parallel local or cloud subagents; conductor runs via Warp's servers ⚠ Parallel agent blocks; no conductor ⚠ Parallel agents in Git worktrees; no live conductor ⚠ Parallel agents via 3rd-party plugins; no built-in conductor
Persistent shared memory across agents & sessions
One local brain Claude · Codex · Gemini all read/write; auto-indexes past chats + the code of repos you open; semantic recall + knowledge graph; encrypted at rest
✅ Local, cross-agent — distills lessons (fixes, decisions, gotchas) from your sessions ⚠ Agent Memory: cloud, preview-only ❌ Per-chat history only ⚠ Manual rules files, no auto-memory ❌ Terminal only — no AI memory
Wire-level token compression before the model
Shrinking the file reads, command output and images the agent sends on every turn — not just showing you how much context is left
✅ Local proxy for Claude Code, on by default — 61% of compressible tool text removed in a replay measured before v1.49, elided text retrievable on demand ⚠ Context-window meter + auto-summarizes history when full; no per-turn compression ❌ None published ⚠ Shows context-window and token usage; no compression ❌ Terminal only — bounded buffer reads via a community MCP plugin
Prompts watched for secrets (97 rules)
Detection, not prevention — the hit is named in a local audit log (DB_PASSWORD, never the value) so you know what to rotate, and a warning banner names it in the app right away
✅ Local, can't be turned off — Claude Code, Codex and Gemini sessions started from Termpolis's launchers ⚠ Secret Redaction (~20 patterns, opt-in) ❌ ❌ ❌
Secret scan at the git boundary
Blocks the commit or push when a key is in the staged diff or an unpushed patch. Real pre-commit / pre-push hooks, installed per repository from Settings → AI Security, cover git typed into any terminal — and keep working with Termpolis closed
✅ Blocks — doesn't just warn ❌ Redaction is prompt-scoped, not git ❌ ❌ ❌
Secrets scrubbed before they enter the AI memory
Stripped before a memory is hashed, embedded, or written — so a key in an old transcript can't be recalled back into an agent later
✅ Prevention, at write time ⚠ Agent Memory is cloud-side ❌ No AI memory ❌ Manual rules files ❌ No AI memory
Scans third-party AI artifacts before install
Skills, plugins, slash-commands, subagents, MCP servers — static scan + install gate
✅ Safe Import — 41 rules, red can't install ❌ ❌ ❌ ❌ Plugin store, no artifact scan
Sensitive-file-read audit (.env, *.pem, ~/.aws/credentials)
Audit-log entry when Claude Code autonomously reads a high-risk file
✅ Built-in ❌ ⚠ Wave AI only: reads need approval, key files blocked; no audit log ❌ ❌
Provider ToS drift watcher
Weekly hash-check on Anthropic and Google terms (OpenAI's page blocks automated fetches, so it is not checked). It files an issue against this repo so the claims on this page stay true — it is not an in-app alert to you
✅ Weekly GitHub Action ❌ ❌ ❌ ❌
Local JSONL audit log
Every prompt-secret hit, block, and sensitive-file read stays on disk — with an in-app viewer
✅ Rotated, on-disk only ⚠ Cloud history ⚠ Terminal scrollback on disk; AI chats in memory only ⚠ Air Teams task & run history (cloud) ⚠ Shell history
Cross-AI context handoff
Resume a Claude session inside Codex / Gemini
✅ One click into Codex or Gemini (agy), or back into Claude Code — bracketed-paste safe ❌ ❌ ⚠ Same IDE, no auto-digest ❌
Cross-agent peer review (Second Opinion)
One installed agent read-only double-checks another's last answer — injection-safe
✅ Any installed agent, per terminal ⚠ Factories review agent checks the PR (early access; another model/harness recommended) ❌ ⚠ Fresh-session reviewer checks the diff ❌
Hunk-by-hunk swarm review w/ test gate
Accept/reject each diff hunk before commit
✅ Built-in review ⚠ Interactive Code Review, no test gate ⚠ AI block edits ⚠ Agent review + diff viewer; no test gate ⚠ Via git add -p
Termpolis-side telemetry on you
Lower is better — what the tool itself collects
✅ None by default · crash reports and usage stats are separate opt-ins, both off ⚠ Analytics + AI/console inputs, opt-out (Free plan: required for AI) ⚠ Anonymized telemetry (opt-out) ⚠ Anonymous usage stats, opt-out ⚠ Mixpanel analytics on by default, opt-out
Price for the editor itself
Provider API costs are separate
✅ Free forever ⚠ Free · Build $20 · Max $200 · Business $50/user ✅ Free ⚠ Free preview (macOS/Linux/Windows) · pricing TBD ✅ Free

Spot something out of date? File an issue and we'll fix it. Notes and sources for each row →

Download and code

Install the app or inspect the source

Available for Windows, macOS, and Linux. Apache 2.0 licensed, open source, free.

Source repository

github.com/codedev-david/termpolis

Explore the codebase, track releases, and contribute.

Browse the repo

Windows

One-click NSIS installer; installs per user, no admin rights needed.

Code Signed (SSL.com) Download for Windows

* Windows SmartScreen may show a warning for new software. Click "More info" then "Run anyway" to proceed. Termpolis is digitally signed and safe to install.

Linux

.deb for Debian/Ubuntu, AppImage for everyone else.

Download .deb (Debian/Ubuntu)
Then in the Downloads folder, run: (apt installs the dependencies the package declares. If the .deb is in your home folder, apt may print a notice that the download is performed unsandboxed as root; that is harmless and the install continues. The launcher ships with --no-sandbox already baked in and the postinst refreshes the desktop and icon caches, so the dock icon Just Works.)
AppImage (other distros)

Support the project

Sponsor Termpolis

Termpolis is free, open source, and Apache 2.0 licensed. Building it (including AI token costs for development) takes time and resources. If you find it useful, please consider sponsoring.

❤️ Sponsor on GitHub