Spend less on coding agents. Control what they do.
London, England
Joined May 2026
- Tweets109
- Following1
- Followers164
- Likes38
Context Mode retweeted
We ran the same long Claude Code session on Opus 5.5 twice, with the same task. The first run cost $7.20. The second, routed through Context Mode Gateway with every feature on, cost $2.49.
That is 65.5% lower. Most of the savings came from not sending the model context it did not need, and the gateway charges no fee on tokens.
contextmode.ai/docs/quick-st…
Context Mode retweeted
thinking about sharing the @contextmode pitch deck publicly.
the idea: skip the cat-and-mouse, let it get passed around, and talk directly with any pre-seed VC who's interested.
would you share your deck in public, or keep it private?
your coding agent stack has more than a routing problem.
context saving: @cmprsr, Induction, Headroom and RTK tackle context optimization with different mechanisms.
gateway: @LiteLLM @PortkeyAI @vercel @Cloudflare cover model access, routing and spend controls.
memory: @mem0ai @supermemory @zep_ai preserve knowledge across sessions.
protect/security: @zenitysec @LassoSecurity focus on agent security and policy enforcement.
session continuity: @specstoryai @GoEntire capture coding sessions and support continuing work.
observability: @langfuse and LangSmith by @LangChain trace agent activity and costs.
@contextmode Gateway brings context saving, recall, cross-agent memory, skills, Protect policies and spend visibility into the request path of existing coding agents.
not identical products. overlapping capabilities across six categories.
compare the mechanisms, coverage and trade-offs. we linked the source docs:
contextmode.ai/docs/landscap…
Context Mode retweeted
Context Mode is a free, local, open source plugin for 10 coding agents. The agent writes a short script, the script runs in a local sandbox, and only the answer enters the context window. The full output stays in a SQLite index on your machine, so the agent can search it later.
2.0 is a full rewrite: 9 tools, hooks on all 10 agents, one database per project, and a bench you can run yourself. github.com/mksglu/context-mo…
Context Mode retweeted
Context Mode, a free plugin for coding agents: 709K downloads from npm, 63K GitHub installs in 14 days, 33K GitHub stars.
that's the exact problem we built session resume for in the Context Mode gateway. you pick up where you stopped, even on another machine, and with Recall the session never compacts or forgets in the first place.
Context Mode retweeted
this is the most common complaint about coding agents right now.
you come back to your session, and it immediately hits compaction. half your context is gone and the agent forgets what it was doing.
almost every reply says the same thing: lower the compact threshold. that's not the fix.
these are the only 4 steps you need to stop losing your session (no bs):
1) stop filling the window. most of it isn't your conversation, it's raw tool output. Context Mode keeps 99% of it out of context.
2) search, don't re-read. every output gets indexed locally, so the agent pulls back the one line it needs instead of the whole log.
3) survive the compaction. session state is restored right after it, so the agent picks up where it left off.
4) install once. npm i -g context-mode@next && context-mode setup
free, local, no telemetry. works in Claude Code, Codex, Gemini, Copilot CLI, OpenCode and more. 32.9k stars got us to v2.
Most annoying thing in Claude Code / Codex: I get notified to come back to a session, then immediately hit compaction.
Easiest fix: if it's already near ~750k (or whatever auto-compact threshold), do the compaction first, and then notify me.
cc @thsottiaux @bcherny
step 2 is the one everyone skips.
this is the most common question in AI dev teams right now.
the agent works. it reviews every PR, fixes the red build, answers every @claude. and the bill keeps climbing.
almost every answer says the same thing: switch to a cheaper model. that's not the fix.
these are the only 4 steps you need to cut your CI agent bill (no bs):
1) count your runs. we scanned 3,530 repos with 12k+ stars, 167 run an agent in GitHub Actions, and the busiest fires every 1.9 minutes. you're paying for every one.
2) keep the model. the cost isn't the model, it's the context it re-reads on every run.
3) change one line. ANTHROPIC_BASE_URL=gateway.contextmode.ai. same workflow, same config, no plugin.
4) compare the bill. we paired @claude, PR review and CI fix runs against direct ones. all three already come out cheaper at the same quality, and the ceiling is 60%+.
2
143
An agent run every 1.9 minutes adds up fast. 60%+ of that bill is on the table, and it takes one line.
Your CI agents are overpaying.
They run on every issue, every PR and every red build. On a busy repo that's an agent run every 1.9 minutes, and you pay for each one.
So we ran those same jobs through the Context Mode gateway, same config, paired against the direct run, via:
① @claude fixes on issues and PRs
② fixing a failing CI run
③ automatic review on every PR
All three already come out cheaper and pass our quality bar, and the ceiling on these jobs is 60%+.
One line in your workflow: ANTHROPIC_BASE_URL=gateway.contextmode.ai
Point ANTHROPIC_BASE_URL at gateway.contextmode.ai and your workflow stays exactly as it is.
contextmode.ai/docs
1
25
𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗠𝗼𝗱𝗲 𝗰𝗼𝗺𝗺𝘂𝗻𝗶𝘁𝘆 𝘂𝗽𝗱𝗮𝘁𝗲
v1 got us to 32.9k stars. v2 shipped three releases on npm next since midnight.
① Hooks proven firing live on 8 agents, and doctor fails loudly if one would quietly never run
② Search answers with 5 clean hits instead of a 15 KB wall
③ A Windows bug you reported this morning, fixed on npm by the afternoon
2,396 tests. 0 failing. 99% of tool output out of context.
Free. Local. Forever.
npm i -g context-mode@next && context-mode setup
github.com/mksglu/context-mo…
2
46
Context Mode turns your coding agent into something you can actually afford and trust, via:
① Saving, which folds the tool output your agent never needed, so a long Claude Code session that cost $4.86 now costs $3.04
② Memory, Recall and Skills, which mean your agent stops re-reading what it already knows, never compacts and starts over, and loads a skill only when the task needs it
③ Protect and Analytics, which let you see every token, every dollar and every command your agents run, and block the ones they shouldn't
① is obvious, ② is less obvious but very powerful, and ③ is the game-changer, because the moment agents write real code for you, you need a control plane, not just a cheaper bill.
Same agent, same model, same key, one command, and every plan gets every feature.
Start free with 1,000 requests, no card, and Pro from $20/mo: contextmode.ai/pricing
Benchmarks with raw data: contextmode.ai/docs/benchmar…
𝚗𝚙𝚡 @𝚌𝚘𝚗𝚝𝚎𝚡𝚝-𝚖𝚘𝚍𝚎/𝚌𝚕𝚒
1
30
Pi is one of the best open coding agents out there. Minimal, fast, yours.
Now it gets what it was missing:
① Saving: large tool output folded before the model reads it. Smaller bill.
② Memory: "use pnpm" on Monday, remembered on Thursday.
③ Skills: the right skill loads when the task needs it.
No plugin. No fork. Pi already lets you set a baseUrl, so we use only that.
𝚗𝚙𝚡 @𝚌𝚘𝚗𝚝𝚎𝚡𝚝-𝚖𝚘𝚍𝚎/𝚌𝚕𝚒 𝚜𝚎𝚝𝚞𝚙 --𝚊𝚐𝚎𝚗𝚝 𝚙𝚒
Same Pi. Same model. Same key. Thank you to the Pi team at earendil-works for building it open.
Link below 👇
The most expensive thing in AI coding isn't the model. It's what you make it read.
① Tests. 400 pass, 1 fails. Your agent reads all 401.
② Docs. You wanted one answer. It read the nav, the footer and the cookie banner.
③ Logs. One error line, buried in 10,000.
You pay for every token. Every request, every day, every engineer.
Context Mode sits between your agent and the model. The model gets the failing test, the error, the answer. Nothing else.
Same agent. Same model. Same key. Smaller bill.
𝚗𝚙𝚡 @𝚌𝚘𝚗𝚝𝚎𝚡𝚝-𝚖𝚘𝚍𝚎/𝚌𝚕𝚒
Cheaper agents won't come from cheaper models. They'll come from smarter context.
How it works, with the numbers: contextmode.ai/docs/benchmar…
Start free, 1,000 requests, no card: contextmode.ai
20
Context Mode solves the "too many coding agents" problem via:
① One command (less setup)
② One gateway (less spend)
③ One memory (less repeating yourself)
① is obvious. You use Claude Code, Codex, OpenCode, Pi, T3 Code, Claude Desktop and the ChatGPT app. 𝚗𝚙𝚡 @𝚌𝚘𝚗𝚝𝚎𝚡𝚝-𝚖𝚘𝚍𝚎/𝚌𝚕𝚒 connects all 7. No plugin, no wrapper, no fork. You keep your app, your model and your key.
② is less obvious, but very powerful. Every request goes through the gateway. Large tool output is folded before the model reads it. Long sessions never compact. The right skill loads when the task needs it. Your rules on what agents may read, run and reach live in one place.
③ is a game-changer. Tell Claude Code "we use pnpm" on Monday. On Thursday, Pi runs pnpm add in the same repo. Your agents stop starting from zero.
Your agents will keep changing. Your context shouldn't have to. Excited to build it with you all!
Link below 👇
